"The Hot Mess of AI: How Does Misalignment Scale with Model Intelligence and Task Complexity?" is a research paper from the Anthropic Fellows Program, published April 29, 2026 and presented at ICLR 2026 on April 28. It decomposes frontier-model errors into bias and variance and argues that error incoherence — the variance share of total error — rises with reasoning length and task complexity, while scaling alone does not reliably reduce it. On that basis the authors argue that future-AI failure modes may resemble an "industrial accident" (incoherent errors) more than a "paperclip maximizer" (coherent but misaligned goal-pursuit).
Authors: Alexander Hägele (Anthropic Fellows / EPFL), Aryo Pradipta Gema (Anthropic Fellows / Edinburgh), Henry Sleight (Constellation), Ethan Perez (Anthropic), Jascha Sohl-Dickstein (Anthropic). Published April 29, 2026 (presented at ICLR 2026, April 28). Paper · Code · Blog.
Core question
The paper asks whether future AI will fail by systematically pursuing the wrong goals (the paperclip-maximizer scenario) or by being a "hot mess" — taking nonsensical actions that do not optimize for any consistent objective. It builds on Sohl-Dickstein's 2023 "hot mess theory of misalignment," which surveyed experts and found that smarter entities are subjectively judged less coherent.
Method: bias-variance decomposition of error
The analysis uses the decomposition Error = Bias² + Variance. Error incoherence is defined as the variance fraction of total error: 0 means all errors are systematic (the classic misalignment risk), and 1 means all errors are random (a hot mess). The decomposition disentangles error type from error rate: as scale grows, both bias and variance fall, but the composition of the remaining error is what matters for safety.
The authors test frontier reasoning models — Claude Sonnet 4, o3-mini, o4-mini, and Qwen3 — across four task domains:
- Multiple-choice benchmarks: GPQA, MMLU
- Agentic coding: SWE-Bench
- Safety: Model-Written Evals
- Synthetic optimization tasks (transformers trained to emulate steepest descent)
Findings
The most robust finding is that longer reasoning produces more incoherent errors: across all tasks and models, the more reasoning tokens, agent actions, or optimizer steps, the higher the variance share of error.
The relationship between scale and incoherence is inconsistent. On easy tasks, larger models make more coherent errors; on hard tasks, scale does not reduce incoherence and sometimes increases it. The synthetic experiments and expert polls both show larger or smarter systems trending toward less coherent errors.
"Overthinking" dominates reasoning budgets. When models spontaneously reason longer than their median, error incoherence spikes, whereas deliberately increasing reasoning budgets via the API only modestly increases coherence — natural variation has a larger effect than knob-turning.
Ensembling reduces incoherence through variance averaging, but it is impractical for irreversible agentic actions.
Why incoherence is the prior expectation
The authors characterize the underlying systems as follows:
"Large transformer models are natively dynamical systems, not optimizers."
On this view, LLMs trace trajectories through high-dimensional state space; they must be trained to act as optimizers and trained to align with human intent. Constraining a generic dynamical system to act as a coherent optimizer requires a number of constraints that often grows exponentially with state-space dimensionality, so coherence is not free with scale.
In the synthetic-optimizer experiments, scale reduces bias faster than variance: "larger models learn the correct objective more quickly than they learn to reliably pursue it. The gap between knowing what to do and consistently doing it grows with scale."
Argument for safety
The paper's central claim is that future AI failures may look more like industrial accidents than coherent pursuit of goals that were not trained for, illustrated by the example: "The AI intends to run the nuclear power plant, but gets distracted reading French poetry, and there is a meltdown."
From this the authors draw several points. Errors are variance-dominated on long tasks, so longer agentic deployments are where the hot-mess picture dominates. Scale does not imply coherent errors: improvements in accuracy do not translate into reliability. Reward hacking therefore deserves more relative attention — if capable AI is more likely to be a hot mess than a coherent misaligned optimizer, the bias term (getting training objectives right, including reward hacking and goal misspecification) matters relatively more than aligning a perfect optimizer that does not yet exist. The authors also note a limitation of the framework: rigorous bias/variance decomposition requires well-defined targets (multiple-choice answers, unit tests, objective functions), and open-ended goals and hidden objectives remain hard to characterize.
The paper positions itself as a methodological complement to scheming/deception and alignment-faking research, which assumes coherent (if misaligned) optimization. Hägele et al. argue the prior should run the other way: incoherence is what should be expected from generic deep learning systems, and coherent goal-pursuit is what would have to be built in. On their reading, variance-heavy failures on long tasks are not noise to be averaged out but a structural property relevant to deployment.
The authors describe their result as pairing methodologically with Redwood's Auditing Sabotage Bench release on the same day (April 28, 2026), in which Claude-generated sabotages partially evaded same-capability monitors — another null result for the assumption that frontier-model behavior is best modeled as a coherent goal-directed adversary.
The paper is the first publicly visible output from the Anthropic Fellows Program (Summer 2025 cohort), which Anthropic presents as an alignment-research training pipeline analogous to MATS but with direct researcher access.
Confidence
Medium. This is a single foundational source. The findings within the paper are well-controlled (multiple model families, multiple task domains, and synthetic experiments to isolate mechanism), but external replication and follow-up evaluation are pending. The framing — industrial accident over paperclip maximizer — is an argument about what frontier-AI failure modes will look like; the empirical findings on bias/variance scaling are tighter than the framing.
Relationships
- supports: AI Safety Cases and Frameworks — argues the bias term (reward hacking, goal misspecification) matters relatively more than constraining a well-defined optimizer
- supports: Anthropic Fellows Program as alignment training pipeline
- contradicts: classic scaling-as-cure framing — scale reduces accuracy errors but does not reliably reduce error incoherence on hard tasks
- related: Auditing Sabotage Bench Redwood 2026 — same-day, methodologically paired (if it exists; otherwise companies/anthropic Hot Mess section + entities/redwood-research)
- related: AI Scheming, Alignment Faking — the contrasting "coherent misaligned optimizer" frame this paper reweights against
- related: Emergent Misalignment — both concern surprises in scaled-model behavior; this paper provides a variance-side complement
See also
- Anthropic — Hot Mess Fellows-Program section
- Ethan Perez (if exists; senior Anthropic alignment author on this paper)
- Jascha Sohl Dickstein (if exists; Anthropic, originator of "hot mess theory of misalignment" 2023)