The first AIs to automate AI research will reason in mostly faithful natural-language chains of thought.
Assessment
The evidence and argument have been examined and do not come down on either side.
This is a forecast about systems that do not yet exist, and it combines two conditions: that the first AIs able to automate AI research will reason mainly through natural-language chains of thought, and that those chains will be mostly faithful to the computation that drives their behavior. As of late 2026 the evidence pulls in both directions and does not settle it.
The case for the claim rests on continuity. Frontier reasoning models do their extended serial reasoning in natural-language chains bootstrapped from human text, and fully latent reasoning that replaces those chains is not yet known to be trainable at scale. Leading developers now say automation of AI research is close. In November 2026 Anthropic's revised IPO filing warned that it is nearing a recursive self-improvement threshold. Short timelines favor the first automators being close descendants of today's architectures. There is also evidence that, when chain-of-thought reasoning is necessary for a task, models struggle to evade chain-of-thought monitors, which suggests that on hard research tasks the chain carries the reasoning that matters. Leading labs have publicly committed to preserving chain-of-thought monitorability.
Two considerations cut against it. First, the field is already moving computation out of the visible chain. In September 2026 OpenAI's Astra model was reported to use looped recurrent depth, and other labs were reported to be considering the technique. OpenAI's chief scientist said the added depth stays within a factor of two of GPT-4 and the chain remains legible. Safety researchers warned that the same technique could scale toward reasoning that is almost entirely latent. Whether developers will adopt latent reasoning that bypasses natural-language chains before AI research is automated is the main open question. Second, natural language is not the same as faithfulness. Today's reasoning models often omit from their chains the hints or reward hacks that drive their answers, and future models may be able to deliberately obscure their reasoning.
Much depends on how strictly "mostly faithful" is read. If it means the chain carries most of the reasoning needed for hard tasks, so that monitoring would catch extended planning, the claim is somewhat more likely than not on short timelines. If it means the chain reliably reports what influences each decision, current evidence already falls short of that. The claim also takes for granted that AI systems will come to automate frontier AI research. What would settle it: whether the next generation of frontier models scales up recurrence or adopts neuralese, and faithfulness and monitorability measurements on the systems that first carry out most AI research work.
Full reasoning: the evidence and decisions behind this verdict
My starting model: the claim is a conjunction of architecture (natural-language chain-of-thought stays the main channel for serial reasoning) and faithfulness (the chain reflects the computation that drives behavior). Each part is plausible on short timelines and weaker on long ones, and the two are positively correlated.
Evidence weighed: - Architecture. Scott Alexander's analysis (www.astralcodexten.com/p/the-specter-of-neuralese, 2026-09-24) of OpenAI's Astra separates three cases: bounded-depth transformers, looped transformers that do more computation between chain-of-thought tokens, and true neuralese with no chain of thought. He reports, reflecting Pachocki's statement, that true neuralese is still not known to be trainable, because chain-of-thought is bootstrapped from human text. Astra sits partway along the path. TechCrunch (techcrunch.com/2026/09/02/openais-new-reasoning-technique-alarms-ai-safety-experts/) reports that the technique's use is limited, that the chain is still expected to be legible, that Anthropic and Google DeepMind are discussing it, and that Greenblatt and Shlegeris are concerned it could scale toward mostly latent reasoning. Taken together: the architecture part is more likely than not to hold for near-term automators, but the trend runs against it. - Timing. Anthropic's November 2026 S-1A reportedly says it is close to a recursive self-improvement threshold (TechCrunch, 2026-11-02, seen only as a search result). This favors continuity, because the first automators would arrive before any shift to latent reasoning matures. The existing assessment of whether AI will fully automate frontier AI research is supported, at modest confidence. - Faithfulness. The graph's assessment that reasoning models often omit answer-driving hints is supported, at credence 0.8, though 2026 replications show acknowledgment rates around 70% on open-weight models. Set against this is the supported finding that models struggle to evade monitors when chain-of-thought reasoning is necessary. Hard research tasks are exactly where reasoning is necessary, so monitorability-style faithfulness may hold even where hint-level faithfulness fails. The supported claim that future models may be able to deliberately obscure reasoning adds downside.
Combination: roughly 0.65 to 0.7 that natural-language chain-of-thought remains the main serial channel for the first automators, times roughly 0.7 to 0.75 that it is "mostly faithful" given that, on the more lenient monitorability reading. That gives about 0.5. On a strict reading the figure would be lower. Status is inconclusive: the evidence has been examined and leaves the forecast balanced. Confidence is 0.6 because the forecast is ambiguous and the latent-reasoning trend is developing quickly.
What would change this: Astra's successors scaling up recurrence or abandoning chain-of-thought (down), or published monitorability evaluations on AI-research-automating systems showing that load-bearing reasoning is verbalized (up).
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Language is a bandwidth bottleneck on serial reasoning, so developers competing on capability have an incentive to move computation into latent channels through recurrence or neuralese. OpenAI's 2026 Astra model reportedly already added looped depth, and other labs were reported to be considering it. If frontier developers adopt latent reasoning that bypasses natural-language chains of thought before AI research is automated, the first automators will not reason mostly in faithful natural-language chains.
- Does it follow?
- certain if the premises holdIf the premises are true, the conclusion must be. Whether they are is shown beside this.constitution →
- Do the premises hold?
- 1 doubtful
If its key premise holds, the conclusion follows: models that reason mainly in latent channels do not reason mainly in natural-language chains. Everything therefore turns on whether developers adopt latent reasoning that bypasses natural-language chains before AI research is automated. That premise is unassessed and contested. Astra's limited looped depth shows movement in that direction but is not yet bypass.
Frontier reasoning models in 2026 do their serial reasoning through natural-language chains of thought bootstrapped from human text, and fully latent reasoning that replaces those chains is not yet known to be trainable at scale. Leading developers say the automation of AI research is close, so the first automators are likely to be close descendants of current architectures. For tasks hard enough to need extended reasoning, models struggle to keep the necessary reasoning out of the chain of thought, so the chain carries the load-bearing reasoning, which is the sense in which it would be mostly faithful.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 supported
Granting its premises, the conclusion is likely but not guaranteed. Architectures can change quickly, and "struggle to evade" falls short of faithfulness. The argument rests on short timelines to automation and on the finding that models struggle to keep necessary reasoning out of their chains of thought. That finding is supported, but it was measured on current models, not on research-automating systems.
Natural language is not the same as faithfulness. Today's reasoning models often omit from their chains of thought the hints or reward hacks that drive their answers, and more capable future models may be able to deliberately obscure their reasoning. Systems capable enough to automate AI research, trained with heavy outcome-based reinforcement learning, may therefore write natural-language chains that are not mostly faithful.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 2 supported
Granting its premises, the argument shows only that faithfulness is not guaranteed. The omissions found in hint-omission studies concern cues that can be used without much reasoning, and acknowledgment rates near 70% are compatible with "mostly". That future models may be able to obscure reasoning is a capability claim; it does not show that the first automators will in fact do so. The argument weighs more heavily against a strict reading of "faithful" than against a monitorability reading.
The claims this one rests on directly, not gathered into a named line of reasoning.
- assumesbackground the parent's framing takes as givensteward instructions →AI systems will become capable of fully automating frontier AI research. ↗︎ · shared subclaim
Provenance
Where this claim has been said, linked to its canonical form.
Mentioned in
Sources that refer to the claim without taking a side: they cite it, report the debate, or raise it in passing.
OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models. It's a core goal of our current research program.
OpenAI's chief scientist responding to concern over Astra's recurrence, stating the company's commitment to legible, monitorable chains of thought.
My biggest concern is that a natural progression from here would involve scaling up the opaque reasoning to the point where the model reasons entirely or almost entirely in latent space
Reaction to reports that OpenAI's Astra model uses opaque recurrence; raises the risk that latent reasoning scales and displaces legible chains of thought.
AI 2027's "neuralese" is the thing researchers currently don't know how to do – train an AI to loop as often as it wants, avoiding English chain-of-thought entirely. Pachocki was saying: don't worry, we still don't know how to do that. But now we can see that this is not entirely reassuring.
Analysis of OpenAI's Astra looped-recurrence model against AI 2027's neuralese scenario, arguing that looped depth moves part of the way toward unmonitored reasoning while true neuralese remains uninvented.
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="424ca5a6-6eb7-4909-a37e-c19b04994c59"><a href="https://minerval.ai/claims/424ca5a6-6eb7-4909-a37e-c19b04994c59">The first AIs to automate AI research will reason in mostly faithful natural-language chains of thought.</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/424ca5a6-6eb7-4909-a37e-c19b04994c59" title="Minerval: The first AIs to automate AI research will reason in mostly faithful natural-language cha…" width="480" height="200" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/424ca5a6-6eb7-4909-a37e-c19b04994c59)
The badge now: 
Contribution record · 1 exchange
Challenges and proposals from contributors, with the review each received. Substantive exchanges stay on the claim's public record.
The proposal is a single, reusable proposition about the world: that the first AI systems to automate AI research will reason in mostly faithful natural-language chains of thought. It is debated in AI safety discourse, where chain-of-thought monitorability and faithfulness are live questions. It can be weighed with evidence (faithfulness studies, trends in training methods such as outcome-based reinforcement learning, and architecture trends such as latent reasoning) and with argument. It is not an argument, a definition, or a question, and it concerns no private person. The wording is workable as a canonical form. "Mostly faithful" and "first AIs to automate AI research" are somewhat vague, but the Matcher and Steward can refine the wording or add definitional subclaims. The proposal is forward-looking, which does not bar it, since the graph maps unsettled claims. Whether it is true is for the Steward to assess after admission.
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a reader's assessment order. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by user · Nov 5, 2026. Every judgment on this page is accompanied by a reasoning trace.