Superintelligence should make us better

It is now a serious possibility, not a science fiction premise, that AI systems will come to exceed human ability across many domains. Nobody knows when, or how evenly, or what the path will look like. Progress may stall in places we do not expect and accelerate in others. But the trajectory of the last several years has been steep enough that it would be careless to plan as though it will simply stop. We think it is worth asking, before the answer is forced on us, what we actually want such systems to do for people.

Most public discussion of this question focuses on capability and risk. Will the systems be powerful? Will they be controllable? Both questions matter enormously, and we take them seriously. But there is a third question that gets less attention and that, in the long run, may shape everyday life just as much: when these systems are woven into how people learn, work, decide and communicate, will they leave those people stronger or weaker?

Capability is not the whole story

A system can be extraordinarily capable and still make the people who depend on it worse off. This is not a hypothetical concern; it is a familiar pattern with ordinary technologies. Navigation apps are wonderful, and many people who rely on them have lost the ability to find their way without one. Autocomplete saves time, and it also narrows the range of phrases people reach for. None of this is a reason to reject the tools. It is a reason to notice that usefulness and empowerment are not the same thing.

With AI the stakes of that distinction grow, because the scope grows. A tool that can draft, summarize, analyze, advise and decide is a tool that can stand in for a large share of human thinking. If it stands in carelessly, people may gradually stop exercising judgment they once had. If it stands in well, it could extend what individuals are able to understand and accomplish far beyond what was previously possible. The same underlying capability can go either way. What decides the outcome is design: what the system is optimized for, what it defaults to, and what it assumes about the person on the other side.

Tools that teach and tools that replace

One useful way to think about this is the difference between tools that teach and tools that replace. A replacing tool gives you the answer and moves on. A teaching tool gives you the answer when that is what you need, but it also makes its reasoning visible, invites you to check it, and leaves you a little more able to handle the next problem yourself.

Both have their place. Nobody wants a lecture on tax law when they just need to know a deadline. But across many interactions, the balance matters. A student who uses an AI system to understand a proof is in a very different position from one who uses it to avoid the proof. A developer who reads and questions generated code builds skill; one who pastes it blindly accumulates risk they cannot see. Systems can nudge toward either pattern through small choices: whether they show their work, whether they ask a clarifying question, whether they point out what the person might want to verify.

We do not think the right response is paternalism. People should decide for themselves how much they want to learn and how much they want to delegate. But systems should make the learning path available and easy, rather than quietly optimizing it away because a direct answer scores better on engagement.

The same capability can leave a person more able or less able; what decides the outcome is design.

Agency, honesty and calibration

Underneath the question of teaching versus replacing is a deeper one about agency. As AI systems grow more capable, it becomes tempting to let them make more decisions on our behalf, and in many cases that will be the right call. But agency is not only about who presses the button. It is about whether people understand what is being done in their name, can meaningfully disagree, and can change course. A system that acts on your behalf without giving you the means to understand or override it has not empowered you, however good its choices happen to be.

Honesty is the foundation that makes agency possible. People can only exercise judgment over a system's output if they can trust what it says about itself. That means not inventing facts, not presenting guesses as certainties, and not telling people what they want to hear because it is pleasant. The failure mode sometimes called sycophancy, where a model agrees with the user rather than informing them, is subtle precisely because it feels good. A system that flatters will often be preferred in the moment and will leave people less informed over time.

Calibration is the more demanding version of honesty. It is not enough for a system to avoid lying; it should convey how confident it is, and that confidence should track how often it is actually right. A well-calibrated system lets people allocate their own attention sensibly, checking where checking is warranted and moving quickly where it is not. A poorly calibrated one, even if it is usually correct, trains people either to trust too much or to distrust everything. As systems become more capable than their users in more areas, calibration becomes one of the few ways a person can still reason about whether to rely on them.

The role of smaller labs

The most capable systems will likely be built by a small number of well-resourced organizations. That is a reasonable consequence of how expensive frontier training is, and we do not think it is inherently bad. But it does mean that many of the choices shaping how people experience AI will be made by others: product teams, application builders, researchers working on specific domains and communities.

Smaller labs have a particular role here. We are closer to specific users and specific problems. We can study how systems behave for people in contexts that are not the center of attention. We can build interfaces that emphasize understanding over mere convenience, and evaluations that measure what actually matters to a given group of people. We can also report honestly on what we see, including where widely used systems fall short for the people we serve.

We are clear-eyed about our limits. A small lab cannot set the direction of the field, and we do not pretend otherwise. But the field is not only its frontier. How these systems land in people's lives depends on a great many decisions made far from the largest training runs, and we want to make ours carefully.

Languages and communities left behind

One place where the question of empowerment becomes concrete is language. Today's AI systems are, in general, much stronger in a handful of widely spoken languages than in the thousands of others. Albanian is a good example, and one we know well. Albanian is personal to us, and one of our offices is in Prishtina. Albanian speakers can use the same systems as everyone else, but the experience is often noticeably worse: less fluent writing, weaker understanding of context, more errors that a native speaker would never make.

This matters beyond convenience. If increasingly capable systems become central to education, work and public life, then people who speak less-served languages face a choice between using a weaker tool in their own language and a stronger one in someone else's. Over time, that pressure could erode the use of those languages in exactly the domains where they most need to remain alive. A future in which AI makes people more capable should not be one in which it makes them less able to think and work in their own language.

Closing that gap is partly a technical problem of data, modeling and evaluation, and it is one we are working on. It is also a question of attention and priority. Communities are not left behind by a deliberate decision; they are left behind by default, because nobody chose to focus on them. We think labs with roots in those communities have both a responsibility and an advantage in changing that default.

What we commit to

We are uncertain about a great deal. We do not know how quickly capabilities will advance, which risks will prove most serious, or how society will adapt. We take the risks of highly capable AI seriously, including the risks of misuse, of systems pursuing goals we did not intend, and of power concentrating in ways that are hard to reverse. We also think it would be a mistake to let fear crowd out the question of what good outcomes would actually look like. Both deserve careful thought.

Within the scope of what we build, we can make some commitments. We will favor designs that help people understand, not just designs that produce answers. We will treat honesty and calibration as core qualities of a system, not as optional polish, and we will try to show uncertainty rather than hide it. We will preserve people's ability to see what a system is doing and to change their minds. We will measure how our products and models serve people in languages and communities that are usually an afterthought, starting with Albanian. And we will be honest about what our own work can and cannot do.

None of this guarantees a good outcome. But we think the question of whether intelligence makes us better is not one that will be answered by capability alone. It will be answered in thousands of choices about how these systems are built and offered to people. We intend to make our share of those choices with care, and to keep asking, as the systems grow more capable, whether the people using them are growing too.