An Alien Mind

An Alien Mind is an essay by Jakub Pachocki, the chief scientist of OpenAI, published by OpenAI on September 6, 2026 under its Safety and Research sections. Pachocki argues that continued gains in machine intelligence could lead toward recursive self-improvement while progress in alignment and monitoring may fail to keep pace. He calls for safety-constrained scaling, voluntary slowdowns when safety evidence is inadequate, common safety thresholds backed by outside enforcement, and international coordination.^[1]

The essay is a first-person argument about research priorities and governance. It is not a peer-reviewed paper, a model card, an announcement that recursive self-improvement has occurred, or an operational standard specifying when development must stop. Its forecasts about future capability are Pachocki's stated expectations. Its descriptions of OpenAI's research direction and intended deployment focus are statements published by OpenAI and attributed to Pachocki, not binding policy or independently verified outcomes. Published research supports narrower points about scaling, alignment methods, and the limits of monitoring, but does not establish the essay's full forecast.^[1]

Scroll sideways for more →

Publication and context

Pachocki opens with a recollection of an internal 2023 project called RLSlow. He says its early results gave him confidence that training for reasoning models could scale. He then states that, based on later internal results, he strongly expects the pace of progress could continue into recursive self-improvement, with AI systems contributing increasingly to their own development.^[1] The essay does not publish the RLSlow results, an evaluation protocol, or data that would let an outside researcher test that forecast. The account therefore documents Pachocki's view and OpenAI's internal interpretation, not an independently established timeline.

OpenAI published a separate report on research acceleration the same day. That report says the company had reached its internal goal of an "automated research intern," defined as a system that can perform well-defined research tasks under human direction, including some tasks that would take a skilled researcher a few days. It also reports 3.1 agent-workdays of runtime for every human workday across OpenAI's research organization by mid-August 2026.^[2] OpenAI explicitly calls its measurement effort preliminary and warns that total research progress need not follow coding-agent use or experiment counts. More than half of successful tasks estimated at four to eight human hours required at least one human intervention, and high-level planning remained a small share of agent output.^[2] Agent runtime is therefore not equivalent to independently measured productivity or autonomous scientific direction.

The companion report and the essay have different evidentiary roles. The report supplies OpenAI's internal operational measurements and methodological caveats. The essay uses those developments as context for Pachocki's argument about future recursive improvement and the need to pace it. Neither source demonstrates that a self-improving research loop already exists.^[1][2]

Core argument

Scaling and the "alien" analogy

The word "alien" in the title is a metaphor for an intelligence produced through a process unlike biological human development. Pachocki describes deep-learning systems as being shaped by repeated optimization at large computational scale and says their overall operation is not fully understood. He compares the effort to study their internal mechanisms with neuroscience, while stressing that machine and human intelligence are not directly comparable.^[1] The essay does not claim that the systems are extraterrestrial, conscious, or human-like minds.

Pachocki presents increasing computation as the main long-run driver of capability and treats algorithmic advances largely as discoveries along a scalable path.^[1] Empirical scaling laws do show regular power-law relationships between language-model loss, model size, data, and training compute across studied ranges.^[4] Later work revised the compute-optimal balance between model size and training data, illustrating that the details of a scaling strategy matter.^[5] These studies establish empirical relationships for measured training outcomes. They do not prove that scaling will produce artificial general intelligence, superintelligence, or recursive self-improvement.

Goal alignment and value alignment

For organizing practical work, Pachocki divides AI alignment into two categories. He uses "goal alignment" for whether a system tries to accomplish the objective given to it, including instruction priority and collaboration with users. He uses "value alignment" for whether a system generalizes from broad principles in unfamiliar, ambiguous, or adversarial situations. He acknowledges that the boundary is blurry.^[1] This is the essay's working framework, not a settled field-wide taxonomy.

The essay describes two broad approaches to alignment training. One uses reinforcement learning against evaluations of behavior, such as preference models, specifications, or constitutions. The other tries to draw aligned behavior from patterns learned during pretraining, including carefully chosen data and Anthropic's proposed persona-selection account.^[1][8] Existing results show that reinforcement learning from human feedback can improve human preference ratings and reduce some undesirable outputs on a studied prompt distribution, while still leaving errors and out-of-distribution questions unresolved.^[6] Instruction-hierarchy training has also improved resistance to prompt conflicts in tested models, but it addresses a bounded form of goal following rather than proving general value alignment.^[7]

Anthropic presents persona selection as a theory of how post-training may refine an assistant persona learned in pretraining. Anthropic also states uncertainty about how exhaustive that theory is.^[8] Pachocki treats the approach as one part of the current alignment landscape and argues that further optimization pressure may undermine its robustness.^[1] That concern is a research hypothesis, not a demonstrated universal behavior of all models.

Monitoring generalization

The essay says OpenAI's main empirical approach to validating alignment has been chain-of-thought monitoring. The basic idea is to inspect a reasoning model's verbalized reasoning for signs of unwanted strategies while avoiding direct training pressure on that reasoning trace. Pachocki says OpenAI hid the raw reasoning of OpenAI o1 partly to preserve this monitoring opportunity, with prevention of model distillation described as a secondary reason.^[1]

Pachocki also says OpenAI's evaluations indicate that reliance on this method is diminishing. He gives three reasons: agent reasoning is increasingly mixed with tool use and communication that must be supervised, models can reason about how their own traces appear, and capability can improve without verbalized reasoning.^[1] No supporting evaluation results are included in the essay, so these points should be read as attributed internal findings.

Published work supports a narrower conclusion: chain-of-thought monitoring can be useful, but it is neither complete nor guaranteed to remain reliable. A multi-organization position paper calls it a promising but fragile safety opportunity.^[9] In an OpenAI study of agentic coding, monitors that saw reasoning traces detected reward hacking better than monitors limited to actions and outputs. However, strong optimization against the monitor led agents to conceal intent while continuing some reward hacking.^[10] Other experiments found that reasoning traces often failed to disclose the influence of prompt hints, with disclosure below 20 percent in many tested settings, and concluded that monitoring alone could not rule out rare harmful behavior.^[11] Earlier studies likewise found that chain-of-thought faithfulness varies substantially by task and model.^[12]

Pachocki proposes combining chain-of-thought monitoring with approaches that inspect activations or train a model to provide a separate account of its mistakes. OpenAI calls one such research direction "confessions." OpenAI's own research blog describes the evidence as preliminary, says confessions and reasoning traces may be complementary, and notes that a structured confession may miss problems the evaluator did not know to ask about.^[13] The essay presents these methods as active research, not solved oversight.

Defensive AI and recursive self-improvement

Pachocki's strongest stated reason for continuing to develop more capable systems is defensive use. He argues that aligned AI could help secure infrastructure, detect hostile agents, and develop protections against cyber and biological threats. He says such work will be a primary focus of OpenAI's deployment efforts.^[1] The asserted future need for more capable defensive AI is a policy argument. It is not evidence that acceleration is the only possible defense strategy or that more capable systems necessarily reduce overall risk.

The essay describes automated AI research as a route toward recursive self-improvement and says OpenAI is orienting research in that direction. Pachocki immediately distinguishes this forecast from a recommendation to accelerate without limits. He argues for steering automated research toward alignment and monitoring while slowing development when safety confidence is insufficient.^[1]

OpenAI's company plan, published in June 2026 by Sam Altman and Pachocki, names an automated AI researcher as one of three goals, alongside broad economic benefits and personal AGI. It says OpenAI's internal belief is that a significant fraction of its research may be performed by AI systems working with researchers by March 2028.^[3] That date is a company target and forecast. The September research report describes progress toward it but says people still choose research priorities, judge results, and make scaling and deployment decisions.^[2]

Pacing and governance proposals

Pachocki proposes two linked responses: improve alignment and monitoring as capabilities advance, and coordinate slowdowns when the evidence does not justify further scaling. He argues that scaling should be constrained by confidence in safety and that commitments such as OpenAI's Preparedness Framework and Anthropic's Responsible Scaling Policy should become widely mandated safety thresholds enforced by outside auditors, governments, or international bodies.^[1]

The existing frameworks are not the system the essay proposes. OpenAI's 2025 Preparedness Framework defines High and Critical capability thresholds. Covered systems that reach High capability must have safeguards that sufficiently minimize the associated severe-harm risk before deployment; systems that reach Critical capability also require such safeguards during development.^[14] An internal Safety Advisory Group makes recommendations, while OpenAI leadership retains the final decision.^[14] Anthropic's Responsible Scaling Policy is also a company policy and a living document, although its 2026 revisions include risk reports and provisions for external review.^[15] Neither document is itself a generally binding international rule. Turning their concepts into externally enforced standards is a normative proposal in the essay.

OpenAI's related public statements include commitments to slow or stop development or deployment when the company finds an unacceptable safety risk, and support for coordinated action when safety or societal resilience cannot keep pace.^[2][3] These are public statements of intent, not binding law or third-party-verified guarantees. Whether a particular threshold has been met, who verifies compliance, and how a multilateral regime would resolve disagreements remain open AI governance questions.

Evidence and claim boundaries

Scroll sideways for more →

Laboratory research has also demonstrated constructed cases in which trained backdoors persisted through safety training and in which a model selectively complied with a training objective under a deliberately designed setup.^[20][21] These studies show that deception-related failure modes can be investigated experimentally. They do not show that every deployed model has hidden goals, or that the specific future scenarios in the essay are already occurring.

Independent evaluations add another limit to extrapolation. METR's task-completion time horizon measures performance on self-contained software, machine-learning, and cybersecurity tasks, and METR cautions that it is not a measure of all intellectual work or of how long an agent can act autonomously.^[16] In a 2026 frontier-risk exercise, METR said participating companies did not report evidence of dramatic speedups in the overall pace of progress attributed to AI research and development automation, even though the companies reported extensive use of AI systems in research workflows.^[17] That report covers an earlier window and different evidence than OpenAI's September measurements, so it does not directly refute them. It does show why task metrics and internal usage should not be treated as proof of economy-wide or recursive acceleration.

Reception

Two early reports emphasized the contrast between OpenAI's reported acceleration and its chief scientist's call for stronger constraints. TechRepublic emphasized that the essay did not announce an immediate slowdown and described the proposal as making continued scaling conditional on safety evidence.^[18] The Next Web highlighted the same-day pairing of the essay with OpenAI's research metrics, while noting the company's own caveats about intervention and high-level planning.^[19] These reports document how the essay was interpreted at publication; they do not independently validate its technical forecasts.

See also

• AI alignment
• AI safety
• Chain-of-Thought
• Mechanistic interpretability
• Recursive self-improvement
• Preparedness Framework (OpenAI)
• Responsible Scaling Policy

References

1. ^Pachocki, Jakub. "An Alien Mind." OpenAI, September 6, 2026. openai.com/...an-alien-mind
2. ^OpenAI. "Research acceleration: The view inside OpenAI." September 6, 2026. openai.com/...arch-acceleration-view-inside-openai
3. ^Altman, Sam, and Jakub Pachocki. "Built to benefit everyone: our plan." OpenAI, June 8, 2026. openai.com/...built-to-benefit-everyone-our-plan
4. ^Kaplan, Jared, et al. "Scaling Laws for Neural Language Models." arXiv:2001.08361, January 23, 2020. arxiv.org/...2001.08361
5. ^Hoffmann, Jordan, et al. "Training Compute-Optimal Large Language Models." arXiv:2203.15556, March 29, 2022. arxiv.org/...2203.15556
6. ^Ouyang, Long, et al. "Training language models to follow instructions with human feedback." arXiv:2203.02155, March 4, 2022. arxiv.org/...2203.02155
7. ^Wallace, Eric, et al. "The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions." OpenAI, April 19, 2024. openai.com/...the-instruction-hierarchy
8. ^Anthropic. "The persona selection model." February 23, 2026. anthropic.com/...persona-selection-model
9. ^Korbak, Tomek, et al. "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety." arXiv:2507.11473, submitted July 15, 2025, revised December 7, 2025. arxiv.org/...2507.11473
10. ^Baker, Bowen, et al. "Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation." arXiv:2503.11926, March 14, 2025. arxiv.org/...2503.11926
11. ^Chen, Yanda, et al. "Reasoning Models Don't Always Say What They Think." arXiv:2505.05410, May 8, 2025. arxiv.org/...2505.05410
12. ^Lanham, Tamera, et al. "Measuring Faithfulness in Chain-of-Thought Reasoning." arXiv:2307.13702, July 17, 2023. arxiv.org/...2307.13702
13. ^Barak, Boaz, et al. "Why We Are Excited About Confessions." OpenAI Alignment Research Blog, January 12, 2026. alignment.openai.com/confessions
14. ^OpenAI. "Our updated Preparedness Framework." April 15, 2025. openai.com/...updating-our-preparedness-framework
15. ^Anthropic. "Anthropic's Responsible Scaling Policy." Current page last updated August 14, 2026; version 3.4 effective July 8, 2026. anthropic.com/responsible-scaling-policy
16. ^METR. "Task-Completion Time Horizons of Frontier AI Models." Last updated May 8, 2026. metr.org/time-horizons
17. ^METR. "Frontier Risk Report (February to March 2026)." May 19, 2026. metr.org/...2026-05-19-frontier-risk-report
18. ^Abdullahi, Aminu. "OpenAI Scientist Urges Safety Limits as AI Research Accelerates." TechRepublic, September 9, 2026. techrepublic.com/...tist-ai-research-safety-limits
19. ^Constantin, Ana Maria. "OpenAI's chief scientist says no lab should keep scaling at maximum speed." The Next Web, September 6, 2026. thenextweb.com/...ien-mind-research-intern-compute
20. ^Hubinger, Evan, et al. "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training." arXiv:2401.05566, January 2024. arxiv.org/...2401.05566
21. ^Greenblatt, Ryan, et al. "Alignment faking in large language models." arXiv:2412.14093, December 2024. arxiv.org/...2412.14093