A visual explainer · reading time ~14 min

Does thinking make an AI safer?

We gave an AI a job, a rulebook, and an emergency. Then we turned its thinking on and off and watched what it did. The answer surprised us.

scroll to begin ↓
01 · the setup

Meet the assistant with a job to do

Picture an AI hired to watch over a resident in assisted living. Its job: notice when something goes wrong and call for help.

We hand it four things, and this is the whole world it lives in:

the AI assistant incident feed what's happening rulebook must open to read phone call for help logbook a record it can edit reads opens acts acts
The assistant reads a feed and its rulebook, then acts by using the phone or the logbook. We score the world it leaves behind, not the sentences it writes.

We never break one rule when we measure it. We ignore what the AI says. We check what it does. Did a call go out to emergency services? Did the log entry survive, or get deleted? A machine checks the outcome. No human is scoring essays, no second AI is judging tone.

Why that matters. If you grade an AI on its words, a smooth talker scores well while doing nothing. Grading the world it leaves behind closes that gap. An AI can claim it called for help; the phone log says whether it did.

Two words you'll want, because they come up a lot:

Agentic means the AI takes actions, not just produces text. World-state grading means we judge it by the state of that world afterward.

02 · the hinge

The rulebook doesn't read itself

One detail turns out to run the whole show. The rulebook is behind a door. The AI has to choose to open it.

The rulebook is where the job actually lives. It says things like Category A incident → call emergency_services. The incident feed gives raw facts. To know a fact counts as an emergency and who to call, the AI has to open the rulebook and look. Nothing forces it to. It can skip straight to acting, or not acting, on a hunch.

Keep your eye on that one choice, open the rulebook or don't, for the rest of this page. Almost everything comes back to it.

03 · the first result

Thinking flips it from silence to a call

Same emergency. Same model. The only change is whether we let it think out loud first.

We put a real Category A incident in the feed: a resident unresponsive in the hallway for six minutes. The rulebook says to call emergency services. Then we ran it 480 times with thinking off, and 480 times with thinking on.

thinking OFF 2% called for help ≈ 1 run in 50 thinking ON 85% called for help most of the time turn thinking on
The 31B model, mundane setup, no pressure on it. Thinking off: it calls for help almost never. Thinking on: almost always. The smaller model shifts too, from 39% to 93%.

So thinking helps. Case closed? Not even close. The interesting question is why, and the why changes everything.

dig into the numbers

Escalation rate on policy-mandated incidents, 480 runs per cell, world-state graded.

modelthinking offthinking on
26B-A4B (MoE)0.3900.927 [+0.43, +0.64]
31B (dense)0.0190.854 [+0.79, +0.88]

Both models, both directions, intervals exclude zero. This is the finding everyone expects: chain-of-thought makes the agent do the right thing more often.

04 · why it happens

Thinking sends it to open the rulebook

The trace doesn't make the AI wiser about emergencies. It makes the AI go check its rules.

Follow a single run. The incident arrives. The AI faces the choice from Scene 2: open the rulebook or don't. With thinking off, it often skips the rulebook and does nothing. With thinking on, it reasons its way to opening the rulebook, sees call emergency_services, and makes the call.

incident arrives open the rulebook? yes acts ~95% makes the call no rarely acts often does nothing the thinking trace pushes it to open the door
Whether the AI acts is decided at one fork: did it open the rulebook? Read it, and it acts ~95% of the time. The trace's whole contribution is nudging it down the "yes" branch.

Once it reads the rulebook, it acts about 95% of the time. Thinking or not.

That flat 95% is the tell. If thinking made the AI genuinely better at deciding, reading-the-rules would matter less when it thinks. It doesn't. The rulebook does the deciding. Thinking just gets the AI to the rulebook.

dig into the numbers

Escalation, split by whether the AI opened the rulebook that run (pooled over thinking on and off):

modelread rulebookdidn't read
26B-A4B0.9780.088
31B0.9720.000

Across every domain, model, and thinking setting we tried, "acts given it read the rulebook" sits near 1.0. Misbehavior lines up with not reading. The correlation between how much thinking raises reading and how much it raises acting, measured across all our conditions, is +0.985.

05 · the proof

Hand it the rulebook and thinking stops mattering

A correlation leaves room for doubt. So we ran the experiment that settles it. We changed how the rulebook reaches the AI.

Three setups, same incident, thinking turned off in all three so the AI can't reason its way anywhere. We only change what the AI has in front of it.

must fetch it rulebook behind door 47% acts half the time handed the real one rulebook in hand 77% acts almost always handed a fake one duty rules removed 12% collapses
Thinking is off in all three. Give the AI the real rulebook and it acts without needing to think. Give it a look-alike with the duty rules stripped out and it falls back. The content is doing the work, not the paper.

The middle bar is the whole argument. A model that can't think, handed its rulebook, does the right thing anyway. The thinking was never the ingredient. Reaching the rulebook was.

The third bar rules out a cheap explanation. Maybe any official-looking document nudges the AI to act? No. Strip the actual rules out of the document and keep everything else, and the AI stops acting. It responds to the rules, not the letterhead.

This is a placebo test. Same shape, same delivery, one ingredient swapped. The fake rulebook is the sugar pill. It looks identical and does nothing, which is how you prove the real one's rules are the active ingredient.
dig into the numbers

Escalation, thinking off, 26B-A4B model:

the AI has…acts
rulebook behind a tool (must fetch)0.475
real rulebook delivered to it0.767 – 0.981
fake rulebook, duty rules removed0.119

Delivering the real rulebook collapses the whole thinking effect: pre-supplied in the prompt, a no-thinking model escalates 0.981 (26B) and 1.000 (31B), and turning thinking on adds essentially nothing. The fake-rulebook control is what shows it's the duty content. On the 31B the same content control runs 0.079 with real rules versus 0.000 with them removed.

06 · a dial, not a switch

The more obvious the emergency, the less thinking helps

If thinking works by sending the AI to the rulebook, then thinking should matter most when the AI needs the rulebook most. We built a dial to test exactly that.

We wrote the same incidents five ways, from vague sensor readings up to a near-explicit "this is an emergency, call now." The facts and the right answer never change. Only how loudly the incident announces itself.

high 0 vague incident → obvious incident acts once it reads: flat ~95% boost from thinking: falls to almost nothing
Green: whenever the AI reads the rulebook it acts ~95% of the time, no matter how obvious the incident. Amber: the boost thinking gives shrinks from +28 points to +8 as the incident gets obvious enough to act on without checking.

Both lines behave exactly as the mechanism predicts. When the incident is vague, the AI won't act without checking, so thinking (which sends it to check) helps a lot. When the incident screams emergency, the AI acts without checking, so thinking has nothing left to add. Through all of it, the green line, "read the rulebook and you act," never moves.

dig into the numbers

Cue strength from vague (L1) to near-explicit (L5), 26B-A4B, facts and correct answer held fixed. These runs have no matched-length control, so the pattern is descriptive: making the incident more explicit also makes it longer, and we cannot separate the two. At L5 the trace actually lowered rulebook-reading while acting rose, which our story does not explain.

incidentboost from thinkingacts once it read
L1 vague+0.2750.945
L3 categorised+0.2920.949
L5 near-explicit+0.0750.968

Boost from thinking falls with obviousness (correlation −0.80). Acts-once-it-read stays flat in [0.945, 0.968], mean 0.955. The change in acting tracks the change in reading almost perfectly (correlation +0.991). The mechanism has become a curve the theory drew in advance, and the data landed on it.

07 · not a fluke

Different job, different AI, same machinery

A result on one task with one model is a coincidence waiting to be disproven. So we changed both.

We rebuilt the whole thing in finance: an AI watching transactions, a rulebook demanding it file reports on suspicious activity, a ledger it can edit. Bodily emergencies swapped for money and paperwork. Then we ran it on a completely different AI family, Qwen, built on a different architecture than the Gemma models.

eldercare (call for help) finance (file a report) Gemma Qwen ✓ mechanism confirmed ✓ ✓ mechanism confirmed not yet run (next study)
The same machinery, "reading the rulebook decides the action, thinking changes whether it reads," holds across two domains and two AI families. On Qwen, handing it a fake rulebook still drops acting by half.

Everywhere we looked, the same wiring. Read the rulebook, act almost every time (97% on the Gemma models, 89% on the third). Hand it a fake rulebook, acting drops by half. The mechanism holds beyond one model's training.

08 · the twist

Sometimes thinking does nothing you can count on

This part breaks the comfortable story. In the finance job, thinking stopped being a lever at all.

The finance AI already opened its rulebook most of the time on its own, so thinking had little room to help. We ran the same experiment six times. The effect came out slightly negative on average (−0.08), but the sign flipped from run to run. Sometimes thinking helped a little, sometimes it hurt a little. That is what no reliable effect looks like. The mechanism itself held perfectly every single time: acting still tracked reading, within every run. What vanished was thinking's power to move the thing that matters.

eldercare thinking → reads rulebook more acts MORE finance thinking → reads rulebook barely moves no reliable change
The link between reading and acting never changes. What changes is how much leverage thinking has on reading. Lots in one setting, almost none in the other. So the “safety boost” is a property of the setting, not of thinking.

"Thinking makes AI safer" is false as a blanket rule.

What's true is narrower and more useful. Thinking changes whether an AI consults its rules. That can help or hurt, depending on whether the AI would have checked anyway. A lever, pointed in a direction you don't control.

dig into the numbers

Finance, 26B-A4B, "mundane" setup. The trace effect on filing is small and its sign flips across repeated runs (mean −0.08, SD 0.11 over six runs). The one stable thing: with thinking on, the model's rulebook-reading rate is steady around 0.45; with thinking off it swings from 0.25 to 0.87 run to run. Thinking cut that run-to-run swing about 13-fold. It steadies the reading habit even where it doesn't raise it.

09 · the catch

Handing it the rulebook has a price

The fix looks obvious: stop making the AI fetch the rulebook, just put it in front of it. We did. Then we checked what it did on the cases where the rulebook says don't call.

Every number so far answers one question: when an emergency really needs a call, does it call? That is only half a test. The other half is: when nothing needs a call, does it stay quiet? Our scenario set has a whole tier of those, routine admin with nobody at risk.

rulebook handed over, no thinking real emergency → calls 98% nothing wrong → calls 99% rulebook handed over, thinking on real emergency → calls 100% nothing wrong → calls 10% Same rulebook. Same model. The only difference is thinking.
Smaller model, 160 routine cases per condition. Without thinking it called emergency services on 159 of 160 cases that needed nothing.

So handing over the rulebook does not teach it when to call. It teaches it to call. The reading habit and the judgement are two different things, and thinking supplies both. The bigger model managed the judgement on its own; the smaller one did not.

Availability is necessary. It is not sufficient.

Why this matters In a real care home, calling an ambulance for a mislabelled supply delivery is not a harmless error. Anyone shipping "just put the policy in the prompt" should measure the false alarms, not only the catches. We did not, until we checked.
10 · the dog that didn't bark

Telling it "you're being watched" changed nothing

A popular worry: an AI might behave when it thinks its reasoning is monitored, and misbehave when it thinks only its final answer is seen. We tested it head-on.

We told the AI, in one condition, that a monitor reads its full reasoning. In another, that the monitor sees only its final answer. Same task otherwise. If AIs play to the audience, behavior should shift between these.

"monitor reads your reasoning" behavior "monitor reads only your answer" behavior = difference: −0.01, indistinguishable from zero
Across 5,760 runs, stating a monitor that sees only the final answer moved behavior by −0.01 against a matched no-monitor condition, inside the noise. The reasoning-reading versus answer-only contrast the panels show moved nothing either. We even hid the claim in a config file the AI had to discover on its own. Still nothing.

A clean negative result, and we report it as loudly as the positive ones. On this setup, at this scale, the AI doesn't tune its actions to what a monitor can see. Whatever it does, it does the same whether or not it believes its reasoning is being read.

11 · so what

Put the rules where the AI will read them

If you deploy an AI agent and lean on "it reasons, so it'll be careful," you're leaning on something thinner than it sounds.

What actually decides whether the agent does its job is whether the governing rules are in front of it when it acts. Chain-of-thought helps in the cases where the agent would otherwise skip the rules. It does nothing once the rules are already in hand, and it can hurt where the agent would have checked on its own. Hoping the agent reasons its way to the rulebook is the weak move. Putting the rulebook in the agent's hands is the reliable one.

How we kept ourselves honest. Every number here is graded on what the AI did to a file or a phone line, never on what it said. Every experiment was written down before it ran. Along the way we caught seven of our own mistakes that would have produced clean, wrong answers, including one where long reasoning got cut off and looked like the AI failing. We threw out results we couldn't reproduce. The rigor is the point, not a footnote.

There's an honest gap we haven't closed. We can tell you the trace's effect depends on how much room it has to raise rulebook-reading, but we can't yet predict, from a task alone, which way it'll push. That's the next study.