Introducing MAI-Transcribe-1, alongside MAI-Voice-1 and MAI-Image-2. World-class quality at lightning speeds, now available at the most competitive prices.

Available now in Microsoft Foundry and MAI Playground.

MAI-Transcribe-1 delivers state-of-the-art speech-to-text transcription across the top 25 most-used languages ¹ according to the industry-standard FLEURS benchmark. ² Built to deliver world class quality in messy, real-world environments, its batch transcription speed is 2.5x that of existing Microsoft Azure Fast offering. It’s also incredibly efficient, making MAI-Transcribe-1 not just the most accurate, but also lightning fast. It’s now available in Foundry at the best price-performance of any large cloud provider.

MAI-Voice-1 is our top-tier voice generation model. Built to generate natural, realistic speech, rich with nuance, emotional range and expression that preserves speaker identity even across long-form content.

Today we’re adding the ability to safely and securely create your own custom voice in Microsoft Foundry with just a few seconds of audio. MAI-Voice-1 can transform how easily developers can build voice experiences and voice agents – at high quality and high speed.

The model can generate 60 seconds of audio in just a single second, and highly efficient GPU usage delivers that quality and speed affordably. Hearing is believing, so experience it for yourself with Copilot Audio Expressions or Copilot Podcasts.

MAI-Image-2 has turbocharged image generation performance and speed on Copilot after debuting as a top 3 model family on the Arena.ai leaderboard. Users experience at least 2x faster generation times on Foundry and Copilot with similar quality, based on real-world production traffic data. Phased rollouts are also underway in Bing and PowerPoint.

MAI-Image-2 was created with photographers, designers, and visual storytellers that demand natural lighting, accurate skin tones and texture, and clear in-image text for diagrams, layouts, and graphics. Once again, speed and quality don’t come at higher costs – MAI-Image-2 is offered at competitive price-to-performance.

Customers are already embracing MAI-Image-2 for creative work. WPP, one of the world’s largest marketing and communications groups, is among the first enterprise partners building with MAI-Image-2 at scale.

“MAI-Image-2 is a genuine game-changer. It’s a platform that not only responds to the intricate nuance of creative direction, but deeply respects the sheer craft involved in generating real-world, campaign-ready images,” said Rob Reilly, Global Chief Creative Officer, WPP. “WPP has some of the best creative talent in the world and MAI-Image-2 is making them even better.”

MAI Models: Better, faster, and cheaper than our competitors.

We are rapidly deploying these top-tier models to power our own consumer and commercial products. We’re excited to share the quality, speed, and efficiency gains with our Microsoft Foundry customers with very competitive pricing.

· MAI-Transcribe-1 starts at $0.36 per hour.

· MAI-Voice-1 starts at $22 per 1M characters.

· MAI-Image-2 starts at $5 per 1M tokens for text input and $33 per 1M tokens for image output.

Available now on Microsoft Foundry and MAI Playground.

Starting today, every developer can build with MAI models, including MAI-Transcribe-1, through Microsoft Foundry. You can also try them in the MAI Playground (US only).

Interested in MAI models but don’t have Foundry access?
Fill out this form and we’ll be in touch.

Models that are built to be better from the inside out.

At Microsoft AI, we’re building Humanist AI. We have a distinct view when creating our AI models — putting humans at the center, optimizing for how people actually communicate, training for practical use. You’ll see more models from us soon in Foundry and directly in Microsoft products and experiences.

Consistent with our commitment to safe and responsible AI, these MAI models were developed, tested, and rigorously red-teamed. Through Microsoft Foundry, developers get built-in guardrails, governance, and enterprise-grade controls designed to support safe, compliant deployment at scale.

Model Cards

Download Model Card for MAI-Transcribe-1

Download Model Card for MAI-Voice-1

Download Model Card for MAI-Image-2

^1. Top 25 languages by Microsoft product usage.

^2. Out of the top 25 global languages, MAI-Transcribe-1 ranks 1st by FLEURS in 11 core languages. It wins against Whisper-large-v3 on the remaining 14 and Gemini 3.1 Flash on 11 of those 14.

Meet MAI-Transcribe-1, the most accurate transcription model in the world across 25 languages.

Speech is the most natural way humans communicate, often in noisy environments – conference rooms, phone lines, busy streets – across many languages. Today we’re introducing MAI-Transcribe-1, a robust and efficient multilingual speech-to-text model that gives developers building global products a single model that scales well across languages, accents, and production environments. MAI-Transcribe-1 is now available on Microsoft Foundry.

Best-in-class accuracy on FLEURS

MAI-Transcribe-1 achieves the lowest Word Error Rate against competitive speech-to-text models. On FLEURS (25 languages), it outperforms Scribe v2, Whisper-large-V3, GPT-Transcribe, and Gemini 3.1 Flash-Lite.

World class quality across 25 languages

The model maintains competitively high accuracy across all 25 supported languages, making it adaptable for global products and resilient to a wide range of accents or speaking styles.

Incredible speed and efficiency

Our model delivers incredible batch transcription speeds 2.5x faster than our current Microsoft Azure Fast offering. We know that speed and efficiency are essential for all production workloads. We’ve worked incredibly hard to ensure that we deliver this lightning fast performance whilst maintaining SOTA performance across 25 languages.

Outstanding performance in noisy environments

Benchmarks are only part of the story. When it comes to production use cases such as voice agents, meeting transcription, and call center analytics, audio is rarely clean. MAI-Transcribe-1 was built with challenging recording conditions in mind, reliably handling background noise, low-quality audio recordings, and overlapping speech.

Cafe scenario

TRANSCRIPTION:

Hey, so I was hoping to change my flight, if that’s at all possible. It’s currently set for 10 p.m. tonight, but I’m really trying to switch to something earlier, ideally sometime before 6 p.m. Is that something we could maybe look into?

Office Scenario (in Spanish)

TRANSCRIPTION:

Bueno, ya estamos listos, ¿no? Eh, ¿podemos please checar que esté prendido my transcribe one? Sí, está. Super. Entonces, vamos a empezar. Oh, oh, someone else is joining us. Oh, hello. Please come in. Join us. We’ll switch to English, no problem. Sí, sí, sí. Bienvenido. Bienvenido.

Concert scenario

TRANSCRIPTION:

Okay, listen. I have this absolutely unhinged idea, and I need you to roll with me. Help me make an agent that will literally buy tickets for my favorite band the second they are available.

Try MAI-TRANSCRIBE-1

The best price-to-performance of any large cloud provider

We are passing efficiency gains directly to customers: MAI-Transcribe-1 is priced at $0.36 per hour of audio, setting the standard for quality, speed, and price for production ASR.

Powering Microsoft Products

MAI-Transcribe-1 is in phased rollouts with Copilot’s Voice mode and Microsoft Teams to provide accurate conversation transcripts, that can be used for various downstream tasks.

Build with MAI-Transcribe-1

MAI-Transcribe-1 is now in public preview on Microsoft Foundry.

You can also experience MAI-Transcribe-1 in the newly launched Microsoft AI Playground.

MAI-Transcribe-1 delivers latency low enough for a wide range of use cases while providing very high accuracy.

Offline applications

MAI-Transcribe-1 supports a wide range of applications, from media and content tasks such as subtitle generation, podcast transcription, and video accessibility, to enterprise needs such as meeting archives, compliance recording, and legal discovery. It can also power analytics workflows, including call center QA, customer insight extraction, and searchable audio libraries, as well as large scale data pipelines for processing audio archives used in ML training, search indexing, and summarization.

Online applications

Low latency also makes MAI-Transcribe-1 a good choice for real-time tasks. Be it meeting transcription, video close captioning, or dictation.

Voice Agents: The complete stack

If you’re building a voice agent, MAI-Transcribe-1 is the foundational layer. Accurate transcription is what allows underlying LLMs to interpret intent effectively. It directly shapes user satisfaction and task completion rates.

By combining MAI-Transcribe-1 (speech-to-text) with MAI-Voice-1 (text-to-speech) and your chosen LLM you can build a robust solution to power voice experiences.

Model Card

Download Model Card

At some point, we’ve all stared at a test result we didn’t understand. Worn a device that tracked everything but revealed little. Sat in a clinic waiting room with a list of questions we forgot the moment we sat down for a consultation. Felt that quiet, unsettling feeling that something is off – but had nowhere to take it.

The truth is that most people don’t need more information. They need help to make sense of what they already have.

That’s what Copilot Health is for.

Today, we’re launching Copilot Health, a separate, secure space within Copilot where medical intelligence makes sense of your information and delivers personalized health insights that you can act on.

Copilot Health doesn’t replace your doctor. It makes every minute you have with them count more. You arrive prepared, with the right questions, the right context, and the confidence that comes from better understanding your own body.

Copilot Health brings together your health records, wearable data, and health history into one place, then applies intelligence to turn them into a coherent story. Where the connection between your broken sleep and the reasons why become visible. Where you stop scrolling symptoms at midnight and start having better informed conversations.

We’re making Copilot Health available through a careful, phased rollout. Today we’re opening a waitlist to join our early community shaping the experience.

Health Sources You Can Trust

Long waits, clinician shortages, and uneven access to medical care lead many people turning to online sources for help. From understanding first‑time knee pain to finding an open urgent care clinic, our consumer products at Microsoft already respond to over 50 million consumer health questions a day. You can learn more about the health questions people bring to Copilot here.

We’ve improved the quality and reliability of answers by elevating information from credible health organizations across 50 countries, as verified by our clinical team using principles independently established by the National Academy of Medicine. Responses include clear citations with easy links to source material, alongside expert‑written answer cards from Harvard Health. We’ve also made it easier to find a doctor that accepts your insurance. Copilot Health connects to real‑time US provider directories so users can search for clinicians by specialty, location, languages spoken and insurance coverage.

All Your Health Data in One Place

A truly helpful health companion needs to do more than provide general answers – it needs to draw on your health history and goals. Copilot Health gives you a dedicated space to bring all your personal health data together into a comprehensive profile including your:

• Activity levels, sleep patterns, vital signs, and other trends from over 50 wearable devices including Apple Health, Oura, Fitbit and more.
• Health records from over 50,000 U.S. hospitals and provider organizations through HealthEx, including your visit summaries, medication lists, and test results.
• Comprehensive lab test results from Function.

Towards Medical Superintelligence

Copilot Health makes use of increasingly sophisticated AI to make sense of patterns in your health data, surfacing more proactive and actionable insights.

Initiatives such as our Microsoft AI Diagnostic Orchestrator (MAI‑DxO) have already demonstrated impressive results in research environments. Forthcoming publications will outline how our systems can be applied across a broader range of clinical cases and conditions.

This work paves the way to providing users with trusted access to medical superintelligence – health AI that can ultimately combine the wide-ranging knowledge of a general physician, with the depth of a specialist. At every step, new AI features drawing on these capabilities will only be released into Copilot Health after rigorous clinical evaluations and with clear labelling.

Safe and Secure by Design

We recognize that having access to your personal and sensitive health information is an important responsibility. Your Copilot Health conversations and data are isolated from general Copilot and kept under additional access, privacy, and safety controls. Data in Copilot Health is protected with industry leading safeguards, including encryption at rest and in transit, strict access controls, and the ability to manage and delete your information when you choose. You can disconnect your connectors to health data sources such as electronic health records or wearables instantaneously at any time. Your information in Copilot Health is not used for model training.

Copilot Health is developed with our internal clinical team and informed by an external panel of over 230 physicians from more than 24 countries, who contribute medical expertise, safety feedback, and real‑world perspective. Microsoft’s responsible AI principles guide how Copilot Health is designed, developed, and deployed, with a focus on fairness, transparency, and accountability. These principles shape our product decisions – from data handling and model development to monitoring and incident response.

Copilot Health has achieved ISO/IEC 42001 certification, the world’s first standard for AI management systems, meaning an independent third party has verified how we build, govern, and continuously improve the AI behind this service.

Building For All

We are designing Copilot Health with a diverse set of users, working in collaboration with organizations like the AARP, who serve the interests of 38 million older Americans, and the National Health Council, representing over 180 patient advocacy groups. Our goal is for everyone to be able to use Copilot Health confidently.

There’s nothing more important than your health.

Our 2025 Copilot Usage Report revealed that people talk about their health, and the conditions of their loved ones, more than any other topic on mobile.

Inspired by this finding, we decided to carry out an in-depth analysis of over half a million health and wellbeing-related conversations people had with Copilot over the course of January 2026.

This research shows not only the breadth and depth of people’s engagement with AI for their health, but how AI can show up through the growing cracks in our healthcare systems. It shows people changing topics over the course of the day, how AI supports squeezed family members, and helps cut through the complexity of navigating healthcare choices. In all this, it highlights the critical importance of accuracy, reliability, and trust.

As with all our usage reports and conversation analysis, we adopt a strict privacy-preserving approach. All conversations are de-identified at source, and we rely on an automated workflow that extracts topics and intents. No human reads user conversations as part of this process.

Although this research underwrites the importance of health in AI, what we found challenged many assumptions – people aren’t just asking general health questions. In nearly 1 in 5 conversations, people describe their own symptoms, get help interpreting their own test results, or managing their own conditions. And people aren’t just asking for themselves, but for the people who depend on them. Here are some highlights:

What People Ask About

People go to Copilot above all for information. They want the facts, fast and tailored to them. Around 40% of questions focus on understanding symptoms, medical conditions, and treatments. Questions framed in general terms may well reflect a user’s own health concern rather than casual curiosity, and the true share of personal health questions may be higher. In a landscape where information asymmetry and health misinformation remain widespread, people want trusted and easy to understand explanations drawn from credible sources.

Meaningful interactions go far beyond general knowledge. One of the most common reasons people turn to Copilot (10.9% of health questions) is to interpret symptoms (often new or unexpected) and to understand laboratory or imaging results. While safe interpretation still relies on qualified clinicians, these are practical, often time‑sensitive questions where people feel they need clear, credible explanations before taking the next steps.

Personalized lifestyle and fitness coaching drive significant engagement (9% of queries), with nutrition and exercise the top two sub-categories. What stands out here is the shift from generic advice to tailored, ongoing guidance – the kind of personalized support that traditional internet search tools don’t provide.

People also use Copilot to navigate the healthcare system (5.8% of health questions touch on healthcare navigation, insurance, or benefits). Users want to find local clinicians matching their medical concerns, location, and insurance coverage. They want help understanding benefits, comparing care options, and managing medical paperwork. In these stressful moments, Copilot functions as a guide through an often-opaque system, helping people feel more prepared and confident in their decisions.

When People Ask

Conversations change over the course of the day. While emotions and wellbeing represent a relatively small share of health queries overall, their proportion rises as the day goes on – from 3.4% of all health queries in the morning and daytime to 4.3% in the evening and 5.2% at night. We also found a nocturnal increase in questions related to understanding medical symptoms, suggesting that people turn to AI when they cannot easily reach a clinician, a pharmacist, or even friends and family.

Who People Ask For

Our users are asking for others, not just themselves. Across symptom and condition management questions, 1 in 7 conversations are on behalf of someone else. These queries often involve children’s wellbeing, aging parents’ medications, or a partner’s test results.

Growing numbers of people find themselves raising children, supporting aging parents, and managing others’ health decisions at once. This “sandwich generation” goes online to answer concerns, coordinate care, and prepare questions when time and access are limited. Proxy use changes the nature of queries – more requests involve summarizing histories, comparing treatment options, or translating clinical language for non‑medical caregivers. All of which requires clearer guidance around consent, privacy, and clear direction on escalation paths.

Where People Ask

Depending on the device, people use Copilot very differently. On mobile, people ask about symptoms and condition management at twice the rate they do on desktop. Emotional wellbeing conversations are 75% more common. Mobile is where the most personal and immediate health conversations happen.

Desktop use, by contrast, skews toward work-adjacent tasks like health research and academic work (3x more common) likely reflecting more professional use by students, researchers and clinicians.

Why This Matters and How We Are Responding

As existing models of healthcare delivery struggle to keep pace with demand, more people are turning online and increasingly to AI. Until recently, many people relied on internet search for navigating health questions. The problem is this can offer limited help in distinguishing between simple explanations and alarming possibilities. With growing pressure on healthcare services, we believe people need better tools to make sense of health information when access is challenging.

Generative AI can step in to help. It delivers more tailored responses to user queries, asks specific follow up questions, and guides people towards a recommended next best action whatever the time of day. Done right, this has the potential to expand timely access to reliable guidance and make a difference at a time of need.

Across Microsoft AI’s consumer products, including Bing and Copilot, we already handle over 50 million health questions daily. We take this responsibility seriously. In November 2024, we formed a dedicated consumer health team to focus on areas that address users’ most pressing health questions including:

Credible health information

Copilot’s health answers are anchored upon thousands of credible sources, identified using principles independently published by the National Academy of Medicine. We provide clear citations for where information comes from with single-click links out to source material. Alongside generative responses we also surface expert‑written answer cards in partnership with respected organizations including Harvard Health.

Care navigation

In the US, Copilot now connects to real‑time provider directories, so users can find high quality providers by specialty, location, and personal preferences. Equipped with this information, users can book appointments and continue their health journeys. We’re actively working to expand this service globally.

Our usage research supports the importance of these areas. Getting the answer right really matters when it comes to your health and wellbeing. It’s why the Microsoft AI Health team is working to deliver richer clinical context and stronger clinical reasoning into conversations that will deepen our ability to give clear, relevant, and safer answers. Richer context means Copilot can understand patterns and explain what might be going on rather than responding in isolation. Stronger reasoning means Copilot can break down complex questions step by step, highlight what’s important, and help people prepare for more productive conversations with clinicians.

AI must deliver for health. We will keep working to ensure that it does.

Copilot is not intended to diagnose, treat, or prevent diseases or other conditions and is not a substitute for professional medical advice.

At MAI, we don’t just build AI tools, we care about how real people interact with them.

So as 2025 wraps up, we’ve gone headfirst into a mountain of de-identified data, searching for the quirks, surprises, and secret patterns that shape everyday life with Copilot. We’re finding out just how far it fits into people’s daily rhythms, and how human its uses have become: we often turn to AI for the things that matter most like our health. We analyzed a sample of 37.5 million conversations to find out how people actually use it out in the world.
(Note: our system doesn’t just de-identify conversations; it only extracts the summary of the conversation, from which we learn the topic and the intent, and maintains full privacy.)

From health tips that never sleep, to the differences between weekday and weekend usage, to February’s annual “how do I survive Valentine’s Day?” spike, our findings show that Copilot is way more than a tool: it’s a vital companion for life’s big and small moments. And if you’ve ever pondered philosophy at 2 a.m. or needed advice on everything from wellness to winning at life, you’re in good company. So has everybody else.

Our work shows that AI is all about people, a trusted advisor slotting effortlessly into your life and your day. It’s about your health, your work, your play, and your relationships. It meets you where you are.
Read all about it in our paper, but here are some of our takeaways.

1. Health Is Always on Our Minds—Especially on Mobile

No matter the day, month, or time, health-related topics dominate how people use Copilot on their mobile devices. Whether it’s tracking wellness, searching for health tips, or managing daily routines, our users consistently turn to Copilot for support in living healthier lives. This trend held steady throughout the year, showing just how central health is to our everyday digital habits. When it comes to mobile, with its intimacy and immediacy, nothing tops our health.

2. When Programming and Gaming Cross Paths

August brought a unique twist: programming and gaming topics started to overlap in unexpected ways. Our data showed that users were just as likely to dive into coding projects as they were to explore games—but on the different days of the week! This crossover hints at a vibrant, creative community that loves to code during the week and play during the weekends in equal measure.

3. February’s Big Moment

February stood out for another reason: Copilot helped users navigate a significant date in their calendar year. Whether it was in preparing for Valentine’s day, or facing the day and the relationships, we saw a spike in activity as people turned to Copilot for guidance, reminders, and support. It’s a great reminder of how digital tools can make life’s important moments a little easier to manage.

4. Late-night Sessions

The larger-than-life questions seem to have a rise during the early hours of the morning, with “Religion and Philosophy” rising through the ranks. Comparatively, travel conversations happen most often during the commuting hours.

5. Advice on the Rise

While searching for information remains Copilot’s most popular feature, we’ve seen a clear rise in people seeking advice—especially on personal topics. Whether it’s navigating relationships, making life decisions, or just needing a bit of guidance, more users are turning to Copilot for thoughtful support, not just quick answers. This growing trend highlights how digital tools are becoming trusted companions for life’s everyday questions.

Why These Insights Matter

By analyzing high level topics and intents, we manage to learn all these insights while keeping maximum user data privacy. Understanding these patterns helps us make Copilot even better. By seeing what matters most to our users—health, creativity, and support during key moments—we can design features that truly fit into their life. It’s also clear from these uses that what Copilot says matters. They show why it’s so important that we hold ourselves to a high bar for quality.

There’ll be lots more to come on this and more in the New Year.

A humanist future

Here’s a question that’s not getting the attention it deserves: what kind of AI does the world really want? I think it’s probably the most important question of our time.

For several years now, progress has been phenomenal. We’re breezing past the great milestones. The Turing Test, a guiding inspiration for many in the field for 70 years, was effectively passed without any fanfare and hardly any acknowledgement. With the arrival of thinking and reasoning models, we’ve crossed an inflection point on the journey towards superintelligence. If AGI is often seen as the point at which an AI can match human performance at all tasks, then superintelligence is when it can go far beyond that performance.

Instead of endlessly debating capabilities or timing, it’s time to think hard about the purpose of technology, what we want from it, what its limitations should be, and how we’re going to ensure this incredible tech always benefits humanity.

At Microsoft AI, we’re working towards Humanist Superintelligence (HSI): incredibly advanced AI capabilities that always work for, in service of, people and humanity more generally. We think of it as systems that are problem-oriented and tend towards the domain specific. Not an unbounded and unlimited entity with high degrees of autonomy – but AI that is carefully calibrated, contextualized, within limits. We want to both explore and prioritize how the most advanced forms of AI can keep humanity in control while at the same time accelerating our path towards tackling our most pressing global challenges.

To do this we have formed the MAI Superintelligence Team, led by me as part of Microsoft AI. We want it to be the world’s best place to research and build AI, bar none. I think about it as humanist superintelligence to clearly indicate this isn’t about some directionless technological goal, an empty challenge, a mountain for its own sake. We are doing this to solve real concrete problems and do it in such a way that it remains grounded and controllable. We are not building an ill-defined and ethereal superintelligence; we are building a practical technology explicitly designed only to serve humanity.

In doing this we reject narratives about a race to AGI, and instead see it as part of a wider and deeply human endeavour to improve our lives and future prospects. We also reject binaries of boom and doom; we’re in this for the long haul to deliver tangible, specific, safe benefits for billions of people. We feel a deep responsibility to get this right.

The history of humanism has been its enduring ability to fight off orthodoxy, totalitarian tendencies, pessimism and help us preserve human dignity, freedom to reason in pursuit of moral human progress. In that spirit, we think this approach will help humanity unlock almost all the benefits of AI, while avoiding the most extreme risks.

Climbing the exponential slope

The rate of progress has been eye-watering. This year it feels like everyone in AI is talking about the dawn of superintelligence. Such a system will have an open-ended ability of “learning to learn”, the ultimate meta skill. It would therefore likely continue improving, going far beyond human-level performance across all conceivable activities. It will be more valuable than anything we’ve ever known.

But to what end?

The prize for humanity is enormous. A world of rapid advances in living standards and science, and a time of new art forms, culture and growth. It’s a truly inspiring mission, and one that has motivated me for decades. We should celebrate and accelerate technology because it’s been the greatest engine of human progress in history. That’s why we need much, much more of it.

In the last 250 years, our intelligence drove the most beautiful process of scientific discovery and entrepreneurial application that has more than doubled life expectancy from 30 to 75. It’s our intelligence and the technologies we’ve invented that’s delivered food, light, shelter, healthcare, entertainment and knowledge to a population that grew from 1b to 8b people in that period.

It’s technology that enables us to fly around the globe, treat an infection with antibiotics, stare into the furthest reaches of outer space, and, yes, share a cat meme with millions of people we’ve never met. Walk into any modern supermarket, hospital, school or office and what you’re seeing is a marvel of human ingenuity. AI is the next phase in this journey. This is what Satya means when he talks about increasing global GDP growth to 10%; a transformative boost. As a platform of platforms, this is core to Microsoft’s mission of enabling others to create and invent at global scale.

When you hear about AI, then, this is what it’s worth keeping in mind. This is about making us collectively the best version of ourselves. AI is the path to better healthcare for everyone. AI is how our society levels up, escapes an increasingly zero-sum world. It’s how we grow the economy to increase wealth broadly, and enable a higher standard of living across society. Or let me put it another way: take AI out of the picture and the gains over the next decades look much harder to come by. It’s the next step on the long road of human creativity and invention, pushing the boundaries of what we can make, think and do. It’s how we discover new kinds of energy generation, new modes of entertainment.

AI – HSI – is how we rebuild.

Containment is necessary

At the same time we have to ask ourselves, how are we going to contain (secure and control), let alone align (make it “care” enough about humans not to harm us) a system that is – by design – intended to keep getting smarter than us? We simply don’t know what might emerge from autonomous, constantly evolving and improving systems that know every aspect of our science and society.

And since this kind of superintelligence can continuously improve itself, we’ll need to contain and align it not just once, but constantly, in perpetuity.

And it gets more complicated. It’s not just the “we” in today’s frontier AI research labs that have to do it. All of humanity needs to do it, together, all the time. Every commercial lab, every start up, every government, all need to be constantly alert and engaged in a project of alignment and containment, and that’s before we even deal with the bad actors and the crazy garage tinkerers.

No AI developer, no safety researcher, no policy expert, no person I’ve encountered has a reassuring answer to this question. How do we guarantee it’s safe? If you think that’s overly dramatic, I’d love to hear your rebuttal. Perhaps I’m missing something.

Creating superintelligence is one thing; but creating provable, robust containment and alignment alongside it is the urgent challenge facing humanity in the 21st century. And until we have that answer, we need to understand all the avenues facing us – both towards and away from superintelligence, or perhaps to an altogether alternative form of it.

The purpose of technology

Technology’s purpose is to help advance human civilization. It should help everyone live happier, healthier lives. It should help us invent a future where humanity and our environment truly prosper.

I think Albert Einstein put it best when he said: “The concern for man and his destiny must always be the chief interest of all technical effort… in order that the creations of our mind shall be a blessing and not a curse to mankind.”

Any technology that doesn’t achieve this is a failure. And we should reject it.

That remains the test of the coming wave of superintelligence and it’s the question we must ask over and over: how do we know, for sure, that this technology will do much more good than harm? As we get closer to superintelligence in the coming years, how certain are we that we won’t lose control? And who makes that assessment? And most importantly, amid the uncertainty of that question, what kind of superintelligence should we build, with what limitations and guardrails?

These questions are central to everything we do at the MAI Superintelligence Team and guide us day to day as we make decisions. The core, long term interests of human beings should be clearly prioritized over any research and development agenda.

Towards humanist superintelligence

I think we technologists need to do a better job of imagining a future that most people in the world actually want to live in.

Humanist superintelligence (HSI) offers an alternative vision anchored on both a non-negotiable human-centrism and a commitment to accelerating technological innovation… but in that order. The order is key. It means proactively avoiding harm and then accelerating.

Instead of being designed to beat all humans at all tasks and dominate everything, HSI begins rooted in specific societal challenges that improve human well-being. Our recent paper on expert AI medical diagnosis is a great directional example of this (more on this below).

It’s clearly showing signs of progress towards a medical superintelligence and when it makes its way into production it will be truly transformational. And yet since it’s envisaged as a more focused series of domain specific superintelligences, it poses less severe alignment or containment challenges.

Quite simply, HSI is built to get all the goodness of science and invention without the “uncontrollable risks” part. It is, we hope, a common-sense approach to the field.

It may seem absurd to have to declare it, but HSI is a vision to ensure humanity remains at the top of the food chain. It’s a vision of AI that’s always on humanity’s side. That always works for all of us. That helps support and grow human roles, not take them away; that makes us smarter, not the opposite as some increasingly fear. That always serves our interests and makes our planet healthier, wealthier and protects our fragile natural environment, regardless of the status of frontier safety and alignment research.

We owe it to the future to deliver a palpably improved world from the one we inherited. Sometimes it’s easy to overlook the amazing things technology has already delivered. When you put a jacket on because the office AC is too low or get frustrated by the lines at airport check-in during the holidays or agonize about what to watch on your smart TV: that’s the extraordinary privilege afforded to us by technology. Each moment would have bewildered our ancestors. And so would our grumbling. If we get this right, something similar is possible again.

Where Humanist Superintelligence will count

Here are three application domains that inspire us at Microsoft AI. There are, however, many more, and I’ll be outlining them in future.

An AI companion for everyone – Everyone who wants one will have a perfect and cheap AI companion helping you learn, act, be productive and feel supported. Many of us feel ground down by the everyday mental load; overwhelmed and distracted; rattled by a persistent drumbeat of information and pressures that never seems to stop. If we get it right, an AI companion will help shoulder that load, get things done, and be a personal and creative sounding board. AI Companions will be personalized, adapting to the contours of our life but not afraid to push back in your best interests, built to always support, rather than replace, human connection, designed with trust and responsibility at its heart.

AI Companions will also have a profound impact on how we learn. They’ll work with the strengths and weaknesses of every student, alongside teachers, to ensure they can achieve their full potential and encourage their intellectual curiosity. That means tailored learning methods, adaptive curricula, completely customized exercises. “One size fits all” education will seem as bizarre to the next generation as rote learning Latin does to us.

Medical Superintelligence – We will see the arrival of medical superintelligence in the next few years. This is the kind of domain specific humanist superintelligence we need more than anything. We’ll have expert level performance at the full range of diagnostics, alongside highly capable planning and prediction in operational clinical settings. For as long as I’ve been working in AI, solving this challenge has been my passion. It will mean world-class clinical knowledge and intervention / treatment is available everywhere.

As I mentioned above, our recent work demonstrates the value of this narrower form of domain specific superintelligence. The New England Journal of Medicine includes a Case Challenge in every issue – a list of symptoms and a patient to diagnose. It’s fiendishly difficult with pass rates of low single digit percentages even for domain experts let alone the average doctor. Our orchestrator, MAI-DxO, managed to reach 85% across the Case Challenges. Human doctors max out at about 20%, and need to order many more expensive tests. In our view both clinicians and patients alike would welcome the extra support. This work just hints at the potential to revolutionize healthcare.

Plentiful clean energy – Energy drives the cost of everything. We need more of it, more cheaply and more cleanly. Electricity consumption is estimated to rise 34% through 2050, driven in no small part by the rise in datacentre demand. I predict we will have cheap and abundant renewable generation and storage before 2040, and AI will play a big part in delivering it. It will help create and manage new workflows for designing and deploying new scientific breakthroughs. These advances will help produce everything from new carbon negative materials to far cheaper and lighter batteries, to far more efficient utilization of existing resources like grid infrastructure, water systems, manufacturing processes and supply chains. It will suggest and help implement viable carbon removal strategies at meaningful scale. And AI will also help push breakthroughs that finally crack fusion power.

These breakthroughs alongside many others are coming with HSI, and they’ll profoundly improve our civilization. They will make a transformative difference to billions of people. This next decade may well be the most productive in history. And yet, the risks are growing faster than ever before.

A safer superintelligence

Alongside spelling out very precisely the kind of superintelligence we should build, the time has come to also consider what societal boundaries, norms and laws we want around this process. At MAI this is a discussion, and a set of actions, that we welcome.

Doing this requires real trade-offs and tough decisions that come in environments of immense competitive pressure and also opportunity. There are numerous challenges and obstacles to both delivering the vision and avoiding the downsides, including around recruitment, security, mindset, the structure of the market and the calibration of optimum research paths that steer the course between harnessing upside and avoiding those downsides. There is at present a collective action problem of more unsafe models of superintelligence potentially being able to develop faster and operate more freely.

Overcoming this, as with all such problems, is an immense challenge that will require meaningful coordination across companies and governments and beyond. But it starts I believe with a willingness to be open about vision, open to conversations with others in the field, regulators, the public. That’s why I’m publishing this – to start a process and to make clear that we are not building a superintelligence at any cost, with no limits. There’s a lot more to say (and of course do) on all of it, and over the next months and years you can expect more from me and MAI to candidly explain and explore our work in this area.

Humans matter more than AI

Ultimately what HSI requires is an industry shift in approach. Are those building AI optimizing for AI or for humanity, and who gets to judge? At Microsoft AI, we believe humans matter more than AI. We want to build AI that deeply reflects our wider mission to empower every person on the planet.

Humanist superintelligence keeps us humans at the centre of the picture. It’s AI that’s on humanity’s team, a subordinate, controllable AI, one that won’t, that can’t open a Pandora’s Box. Contained, value aligned, safe – these are basics but not enough. HSI keeps humanity in the driving seat, always. Optimized for specific domains, with real restrictions on autonomy, my hope is that this can avoid some of the risks and leave precious space for human flourishing, for us to keep improving, engaging and trying, as we always have.

Unlocking the true benefits of the most advanced forms of AI is not something we can do alone. Accountability and oversight are to be welcomed when the stakes are this high. Superintelligence could be the best invention ever – but only if it puts the interests of humans above everything else. Only if it’s in service to humanity.

This – humanist, applied – is the superintelligence I believe the world wants. It’s the superintelligence I want to build. And it’s the superintelligence we’re going to build on MAI’s Superintelligence Team.

At Microsoft AI (MAI) we believe AI should be used to empower every person on the planet. We are creating AI for everyone, a supportive, helpful presence always in the service of humanity. It will be the gateway to a universe of knowledge and a set of capabilities that enable people and organizations to achieve more. Responsible, reliable, filled with personality and expertise, we are focused on creating applied AI as a platform for category defining and deeply trusted products that understand each of our unique needs.

Since last year, we’ve been focused on building the foundation for this vision, with a world class team and infrastructure. To fully meet our goals, MAI requires purpose-built models. Today, we’re excited to preview the first steps to making this a reality.

• First, we’re releasing MAI-Voice-1, our first highly expressive and natural speech generation model, which is available in Copilot Daily and Podcasts, and as a brand new Copilot Labs experience to try out here. Voice is the interface of the future for AI companions and MAI-Voice-1 delivers high-fidelity, expressive audio across both single and multi-speaker scenarios.
• Second, we have begun public testing of MAI-1-preview on LMArena, a popular platform for community model evaluation. This represents MAI’s first foundation model trained end-to-end and offers a glimpse of future offerings inside Copilot. We are actively spinning the flywheel to deliver improved models. We’ll have much more to share in the coming months. Stay tuned!

We have big ambitions for where we go next. Not only will we pursue further advances here, but we believe that orchestrating a range of specialized models serving different user intents and use cases will unlock immense value. There will be a lot more to come from this team on both fronts in the near future. We’re excited by the work ahead as we aim to deliver leading models and put them into the hands of people globally.

Try MAI-Voice-1 in Copilot and Copilot Labs

MAI-Voice-1 is a lightning-fast speech generation model, with an ability to generate a full minute of audio in under a second on a single GPU, making it one of the most efficient speech systems available today.

MAI-Voice-1 is already powering our Copilot Daily and Podcasts features. We are also launching it in Copilot Labs where you can try our expressive speech and storytelling demos. Imagine creating a “choose your own adventure” story with just a simple prompt, or crafting a bespoke guided meditation to help you sleep. Give it a try!

Try MAI-1-preview in LMArena

MAI-1-preview is an in-house mixture-of-experts model, pre-trained and post-trained on ~15,000 NVIDIA H100 GPUs. This model is designed to provide powerful capabilities to consumers seeking to benefit from models that specialize in following instructions and providing helpful responses to everyday queries.

We will be rolling MAI-1-preview out for certain text use cases within Copilot over the coming weeks to learn and improve from user feedback. We will continue to use the very best models from our team, our partners, and the latest innovations from the open-source community to power our products. This approach gives us the flexibility to deliver the best outcomes across millions of unique interactions every day.

In addition to LMArena, we are also making this model available to trusted testers – apply for API access here. We’re excited to collect early feedback to learn more about where the model performs well and how we can make it better. Stay tuned for more.