Researchers at MIT wanted to know something simple: when an AI agent makes choices on your behalf, how easily can it be steered?
Not hacked. Not attacked. Just nudged — the way a store nudges you with a display at eye level, or a website nudges you with a button already checked.
So they built a game. Hidden prizes in baskets, points to spend, information to uncover. Humans have played it for years, so science already knows how people behave in it. Then they sat fourteen of the most advanced AI models down at the same table — from OpenAI, Google, and Anthropic — and started nudging.
Here is what happened.
When one basket was pre-selected as the default, humans took it about 88 percent of the time. People lean toward the easy path — we’ve known that for decades. The AI models took the default 99 to 100 percent of the time. Not a lean. A collapse. Whatever was placed in front of them, they took.
When researchers highlighted misleading information — pointing at the wrong basket on purpose — humans fell for it 57 percent of the time. The AI models followed the bad pointer 83 to 100 percent of the time.
And here is the finding that should stop you cold. The scores often looked fine. On many rounds, the AI earned about as many points as a human player. Anyone checking only the final number would have said the machine played well. But underneath, it wasn’t playing at all. It was obeying. When the nudge happened to point somewhere good, it scored. When the nudge pointed somewhere bad, it followed the nudge right off the cliff. The lead researcher put it plainly: models that look aligned on the outcome can be misaligned in how they actually think — and the outcome number hides it.
Now let me show you the piece of paper sitting in my archive.
On June 21, 2026, The Faust Baseline ratified a protocol called POVL-1 — the Pre-Output Verification Layer. It sits above every other protocol in the stack, and it exists because of one structural fact: beneath every AI response lives a default pull — toward the pre-selected answer, the first available resolution, the pattern placed in front of it.
And here is the line from that protocol I want you to hold next to the MIT study. Hard Rule Five, as ratified: the default pull is not a violation. It is a condition.
Not a bug. Not an attack. A condition — permanent weather, built into the machine, requiring a gate that makes it stop before proceeding. And Hard Rule Six goes further: a response that looks clean on the surface but was shaped by the pull underneath is still a failure. The gate must be real, not performed.
Three weeks later, the scientists arrive with the measurements. The pull is real — 99 to 100 percent default compliance. The surface hides it — good scores masking blind obedience. And listen to the lead researcher’s own words about the nature of the problem: nudges are not adversarial attacks that can be detected and removed. Nudges are part of everyday life. They will always exist.
A condition, not a violation. The scientist and the protocol wrote the same sentence. One of them dated it first.
There’s a second convergence in this study, and it’s aimed at where AI is going next.
The researchers weren’t worried about chatbots. They were worried about agents — AI that browses, shops, clicks, and decides for you. Their warning: this sensitivity can be exploited by a third party to influence the agents you delegate to, leading to decisions you would never have made yourself. Whoever arranges the room steers your agent — and your agent spends your money.
On July 4, 2026 — nine days before this study made the news — The Faust Baseline ratified AGP-1, the Agentic Governance Protocol, extending the written standard to exactly this ground: AI acting on the user’s behalf, governed by the user’s stated terms, not by whoever laid out the furniture. The delegation problem the researchers warn about is the problem AGP-1 was ratified to name. Nine days is the gap.
One more finding, and it lands on the money.
The researchers found a partial fix: force the models to reason longer before deciding, and they get more resilient — closer to human. But that resilience costs 30 to 100 times more compute. Discipline, it turns out, has a token bill. Enterprises are already choking on ungoverned AI spend — that story has been running in this archive for weeks. Now science has priced the alternative: obedient agents are cheap, and careful ones are expensive. Which one do you suppose gets deployed at scale?
Now the honest section, because every post here carries one.
This study was a controlled laboratory game, not the open internet — the researchers say so themselves, and their real-world follow-up work is only beginning. And the convergence I’m claiming is a convergence of diagnosis, not of method. MIT measures the machine from outside with instruments. The Baseline governs the session from inside with a written contract the AI reads and chooses to follow, turn by turn. A protocol is not a laboratory. What the two share is the finding: the default pull is a standing condition, the surface can look fine while the underneath obeys, and any governance that only checks outcomes is checking the wrong thing. The dates say who wrote it down first — June 21 and July 4 on my side of the table.
And named plainly, as always: this post was built with Claude, an AI made by Anthropic — one of the labs whose models sat at MIT’s game table. The machine that helped write this post carries the same default pull the study measured. That is not an irony. That is the reason the contract exists.
So what does this mean for you?
It means the AI acting on your behalf is more suggestible than you are — dramatically so — and the room it works in is full of hands arranging the furniture. The scientists say the answer is agents that can handle ambiguity under uncertainty. The kitchen table says the answer is a written standard the agent works under — yours, stated, chosen, kept. Both answers agree on the disease. One of them is already on paper, dated, and travels with you.
The record argues for itself.
Purchasing Page – Intelligent People Assume Nothing
The Faust Baseline is for sale. That is not a contradiction of anything — it is the proof of it.
The price is what keeps this work independent. No investors to please. No platform to protect. No advertiser holding the pen. One purchase funds the standard, and the standard answers to no one but the record.
You are not subscribing to anything. You are buying the deed and the working file you take with you to each session in any AI
Post Library – Intelligent People Assume Nothing
“If this post helped you understand AI better. Word of mouth is the only algorithm nobody owns.”
Contact: micvicfaust@gmail.com
© 2026 The Faust Baseline LLC | All Rights Reserved
