We were taught to take the emotion out of decisions. The machine was built that way—which is why committing is still ours. Every real decision still ends in a person.
By now you’ve likely caught up on the call for slowing AI development. ICYMI, Dario Amodei has called for frontier AI companies to slow the pace of capability development so safety work can catch up. He is proposing embedded independent evaluators, industry coordination, and eventually international coordination. Sam Altman and Elon Musk have publicly backed the call, making this a rare moment of alignment among competing frontier labs. But it is worth being clear that this is not an actual pause or halt in training.
We are both visiting researchers at UC Berkeley’s Center for Human-Compatible AI, Stuart Russell’s lab, so we feel very privileged to have some insight into how some of the world’s top safety researchers are thinking about this moment. The short version is that they are much happier now than they were last week.
We take AI safety really seriously. We might not know how AI would kill us all, but we don’t know that it won't. We know that the cyber risks are unacceptably high—especially because the frontier labs' cybersecurity is unacceptably weak. With our background in the grid and finance, we both know how reliant modern life is on a small number of enormously important systems. If slowing down gives the labs more time to secure those systems, that alone seems worthwhile to us.
But our particular perspective is a little different. We think most people want AI that preserves their agency and autonomy. Unsafe AI clearly interferes with that. But there is another way AI can interfere with human agency that doesn’t require anything as dramatic as an AI catastrophe.
It can make it harder for you to remain the author of your own mind and your own work. That is what we care about. And we think authorship depends partly on what the AI itself is like.
You need to be able to keep up with it. You need to be able to tell when it is telling you the truth. You need some visibility into what it has done and how it arrived at what it is telling you. And ultimately you need to be able to make the judgment yourself.
Those sound like slightly different things, but they collapse pretty quickly in practice. If we can’t keep up with a system, we defer to it. If we can’t tell whether it is telling us the truth, we have to trust it. If we can’t see what it has done, we can’t really take responsibility for the result. And if we're no longer making the important judgments, we're not really the author anymore. This is why the frontier safety conversation matters to us even though most people will never work on a frontier model.
Look at where the big labs are heading. First, they are racing toward recursive self-improvement: AI that can help build better AI. Stuart treats recursive self-improvement as a redline. A “do not do.” Amodei’s proposal is unusually explicit about this. His framework includes eventual international limits on the most dangerous capabilities, including, possibly, a “speed limit” on recursive self-improvement, negotiated even with China.
The reason we care about that isn’t simply that AI that makes itself smarter by itself sounds frightening. It is that the faster these systems can change, the harder it becomes for humans to keep up with what they are doing, what they are capable of, and what we should trust them to do and we are already seeing one version of this with reasoning.
Right now, we can see something of how these systems reason. We can read their chains of thought. Forty-one safety researchers recently signed a paper arguing that monitorable reasoning is a real safety signal, and that ordinary development choices could erode it. Neel Nanda at Google DeepMind wrote that the idea that keeping chain of thought monitorable doesn’t matter is, in his words, “total bullshit.” It is one of our best current tools for seeing what a model may be trying to do. Losing it would be a tragedy.
OpenAI’s new Astra model is already harder to monitor. OpenAI’s own system card says Astra’s chain-of-thought monitorability has decreased relative to its predecessor, and that Astra is better able to control what appears in its chain of thought. It can also do more reasoning without producing a corresponding visible reasoning trace.
We are building systems to help us think while simultaneously making it harder to see what they are doing. If we can’t tell whether the system is telling us the truth, what it has actually done, or how it arrived at the answer, we can still use the output but we have a much harder time saying that we authored the work. That’s why this conversation about slowing the frontier matters to our work.
ARIA Institute / Artificiality Institute Summer Fellowship and Design Hackathon
In August we hosted our first fellowship week with graduate students from the ARIA Institute—a week of research, arguing, and prototyping. The fellows came from different disciplines and left with three designs in development: a multi-human/multi-agent collaboration design, a feedback-disclosure design, and a design testing the impact of effortful AI collaboration.
We believe fellowships that translate human research into new designs can create broad impact by showing people that AI can be different—and, ultimately, better for humans. We are currently fundraising to support our ongoing fellowships with ARIA and welcome any ideas, introductions, or support that could help advance this important work.
Thanks to everyone who took the survey that fed the week’s work. That survey deserves its own report-back, so here it is...
We asked people to describe two AI-assisted tasks: one that felt very much like theirs, and one that didn’t. The counterfactuals ran backwards from what you’d predict. Owned work was more AI-dependent: 9 of 37 owned tasks could not have been done without AI, against 4 of 38 disowned ones.
Graphic novels made possible, a children’s book they could now illustrate, the filing automation a technician built to their spec. What they disowned was work they could do easily: emails, LinkedIn posts, a concluding paragraph. Writing that was supposed to sound like them but then came back generic.
The pattern was that ownership followed judgment rather than production. People owned what they revised, questioned, and reshaped, and disowned output they couldn’t evaluate. One student who shipped a working website said she wouldn’t even feel responsible for patching it.
Sequence mattered too. Drafting first and using AI to refine tended to preserve ownership, while handing AI an idea and asking it to generate lost it when the output was meant to carry the person’s voice. Some people found ownership through rejection—a bad draft showing them what they actually wanted. And the heaviest users sometimes located ownership in the systems they’d built—trained agents, curated context—rather than in any single output.
The usual caveats apply, and then some. The sample is small, self-selected, and expert: mostly daily users, many already part of this community. About a third of respondents rejected or subverted the owned/disowned pairing. Some never feel ownership of AI-assisted work; some always do. We’re treating that as a finding about how people construe ownership rather than as missing data. These are patterns the fellowship designs will test, not established results.
Two papers have now come out of the week with the fellows.
The first is the survey work itself, written up and submitted to the main NeurIPS conference in Sydney this December. It’s under anonymous review so we can’t share the paper yet.
The second takes one of the questions raised by the survey and tests it experimentally. It’s headed for a NeurIPS workshop at the Paris satellite the same week. The fellows designed studies around advice-giving—one of the most human things we do with what we know—looking at what happens to the feeling of authorship when AI helps, whether people disclose that help, and how the person receiving the advice responds.
We spent a lot of time talking about disclosure. Whether people are honest about AI use comes down to pride and shame as much as rules, and those emotions are being renegotiated everywhere right now. Watching the fellows build studies that can actually measure some of that was one of the best parts of the week.
We want to extend the fellowship, bringing multidisciplinary teams together each year to shape the future of AI products. We can’t do it alone, and we’d welcome recommendations for individuals and organizations who might support the program.
We're launching something new. For years people have asked us to bring our research to their teams—just get in front of everyone and help them make sense of what AI is doing to their work. So we built exactly that: the Stay Human Briefing. To launch it, we wrote four pieces that walk through the thinking behind it—why people feel the way they do about AI, and what growing with it actually takes.
The first one lays out what we do about all of it with teams: the map of what's irreducibly yours, the three ways AI is changing your thinking, and a daily practice for staying the author of your own mind. The second starts with a strange fact: half of American adults now use AI, and the more they use it, the less they like what it's doing to them—and underneath the complaints is a fear nobody writes down. The third is for leaders: when every competitor rents the same model at the same price, your people become the only advantage left. The fourth follows the money nobody tracks—generation went to nearly free, verification didn't, and the checking turns out to be where judgment forms now.
The Stay Human Briefing is sixty minutes with your team, remote and live. We open with what a decade of research has taught us—why your people feel so differently about AI, what stays irreducibly human, how dependency happens, and what staying the author takes. Then we hand the hour over. The questions people are actually carrying—am I still doing my job, can I trust what I made, how do I grow with this—are what we answer best, live.
Check out more here and get in touch about your Stay Human briefing.
We were all taught to take the emotion out of decisions. Chapter eleven is about the man who actually did. Elliot lost the tissue connecting thinking to feeling, kept all of his intelligence, and could never decide anything again—because deciding finishes with a feeling, and nothing in him ever tipped. Which matters now, because you share your deciding every day with a machine built exactly that way. Nothing in it changes when something matters.
So this chapter is about two techniques you can use—feelings are data about you, evidence is data about the world—and the one feeling you can't take at face value anymore: the feeling of being right. Fluency manufactures it. The machine does it in full sentences, to you personally. From there: when your gut has actually earned your trust versus when it's assuming an authority it never earned, why traders and analysts learn to be wrong so differently, and the one skill in all of this with an actual score—calibration, the small ritual Dave and I have run for five years, and how AI changed it.
We wanted to update our declaration on how we personally use AI.
On working with it: We use AI in nearly everything we do—research, analysis, coding, writing. Much of our work could not exist without it. Every piece carries our names and our full accountability: we made it, we stand behind it, and we can tell you exactly how the machine was involved.
We know how that sits next to what we study. We track what AI erodes in human thinking, and we use it every day. We think you can't do one without the other. Conscious use is the practice we teach, so it has to be the practice we live—in public, on the record, where you can check.
One line we've drawn: since September 2024, we haven't used image or video generators. We can't find authorship in them—too little of the creative act stays in our hands. Others may find it there, and the tools will change. For now, anything visual here was made by a person.
People sometimes run our writing through AI detectors. We understand the anxiety—it belongs to this moment. We're working on a longer timescale: humans and these systems shaping each other for generations.
On what our work is made from: AI is trained on human work, mostly taken without asking. We publish anyway, knowing our words feed the machines too. What we push for is a future where newly human-made work keeps its human value—where the people who write, draw, and think are paid and named, and feeding the machine is a choice, not a condition of being online. How that gets built is an open question. That it must be built is not.
On what it costs: AI runs on data centers, and data centers run on power, water, and land—often someone else's. We come from the energy world; we don't wave this away with efficiency promises, and we don't inflate it either—an AI query costs a fraction of what streaming video does, and honest accounting matters in both directions. But the real choice is architectural. One of our deepest design principles is the edge: AI that is private and efficient, running close to you, over AI that is frontier and greedy, running on ever-larger machines somewhere else. Minds for our minds shouldn't require a power plant each.
On jobs: We know from all of human history that a society of people without purpose is a society in trouble. Work is where most people find a large part of theirs. So we say it explicitly: we want AI that creates more purposeful jobs, and we want the jobs that remain to be more human—more judgment, more craft, more of what only a person can bring—because AI took the rest. Human-enhancing is the test. If a deployment makes work emptier for the people doing it, it fails, whatever it saves.
Note: this declaration is now posted on our About page for all to read.
Chris Messina invented the hashtag. Then he gave it away.
That one fact tells you a lot about him. He helped push Firefox to its first 100 million downloads, co-organized the first BarCamp and spread the unconference model to hundreds of cities, opened San Francisco's first coworking spaces, and designed products at Google and Uber. On Product Hunt he's hunted more products than anyone alive. When Chris pays attention to something, it usually means the rest of us will be paying attention to it soon.
At the Summit he's running a two-part salon called Everyone Can Build Now. Friday you vibe code a real product with him—ideation, design sketches, implementation, launch plan, the whole sequence. Saturday you come back and finish it. You arrive with an idea. You leave with something you built.
And tickets for the Artificiality Summit this October:
Don't Wait: Price is $1,995 (or $2,995 for two).
Don't Worry: We have updated our cancellation policy to better align with similar events and organizations. Through September 22 (30 days prior to the Summit), you can receive a 50% refund. At any time, you can transfer your ticket to someone else or you can apply the full value of your ticket to next year's Summit. You can see all the details on the Summit page.
Don't Forget: We provide a 30% discount for education and nonprofit customers. Send us an email for more information.
Artificiality Summit 2026
Join us October 22-24, 2026 in Bend, Oregon for 2.5 days with a fantastic group of speakers—academics, authors, designers, investors, photographers, and more.
We'll explore the theme of Unknowing—not ignorance, but a necessary release of inherited assumptions. We don’t yet know what AI will become, and we don’t yet know what we will become in relationship to it.
Unknowing is the space between—the place where neither side is fixed, and something new can emerge.
A heads up on our travels: We're planning to be in Amsterdam, Miami, Munich, New York, and Venice (Italy) later this year. Some stops are for keynotes for corporate events...and some stops will be for our soon-to-be-announced Thinking With AI course + certification we are developing. If you've made it this far and would like an early heads up on this, just reply to this email. We've already had quite a few inquires and will be releasing details very soon. If you want to talk about this for your whole team, get in touch!
Follow us on your favorite channels—and please like, share, and repost to help us spread the word.
Helen Edwards is a Co-Founder of Artificiality. She previously co-founded Intelligentsia.ai (acquired by Atlantic Media) and worked at Meridian Energy, Pacific Gas & Electric, Quartz, and Transpower. Contact at helen@artificialityinstitute.org.
Dave Edwards is a Co-Founder of Artificiality. He previously co-founded Intelligentsia.ai (acquired by Atlantic Media) and worked at Apple, CRV, Macromedia, Morgan Stanley, Quartz, and ThinkEquity. Contact at dave@artificialityinstitute.org.
