False Easiness: When LLM Agents Find the Commons Decision Too Simple

Groups of language-model agents that share a resource reliably ruin it. In a recent repeated commons simulation, nearly every episode ended with negative total welfare, yet each agent's choices were close to the best reply to what the others were doing. This proposal argues that the failure is a mismatch in perceived difficulty. Deciding what is good for me this round is easy and feels easy. Deciding what is good for everyone over the episode is hard, because the harm from my extraction builds up slowly, lands on people who never speak, and shows up only after many steps of reasoning. Agents answer the easy question, because nothing in the situation signals that it is the wrong one. We call this false easiness: a decision that feels easy while the computation that actually matters for welfare is deep.

The claim is testable and goes beyond saying that agents are selfish or that prompts matter. We will measure, for every extraction decision, how hard the agent treated the choice (from its reasoning trace and its hidden state before it writes anything). We will also measure how deep the welfare-relevant reasoning objectively is (from the simulator's known dynamics). We predict that over-extraction concentrates where objective depth is high and perceived difficulty is low. We further predict that interventions help to the extent they close that gap. One such intervention is a short, verifiable statement of where the agent's own proposed action leads. That should beat interventions that merely add information.

Where this sits in prior work

Two literatures meet here, and neither has crossed into the other. Work on perceived difficulty divergence asks why formal hardness fails to predict how hard a task feels. In puzzle games, solution length is a poor proxy for difficulty (The Unexpected Consequence of Incremental Design Changes). Generalized Entropy and Solution Information for Measuring Puzzle Difficulty explains part of the gap. It shows that difficulty depends on which inference rules a solver has. Its information measures correlate with human ratings of The Witness puzzles at Pearson $r=0.563$, rising to $0.751$ with a learned policy. Giving the solver model more deductive power lowers the fit, because people miss inferences. We borrow that move directly. We define objective difficulty relative to a family of bounded solvers rather than an ideal one.

Two recent papers measure difficulty as a model experiences it, and we use both as measurement tools. Cognitive Episodes in LLM Reasoning Traces splits reasoning traces into sentence-level episodes such as planning, checking and backtracking. Features built from these episodes predict human test-item difficulty, with a 24.4% RMSE reduction over strong baselines. Harder items produce more iterative reasoning. The LLM Already Knows shows that a model's sense of a question's difficulty can be read linearly from its hidden state before the first token, reaching ROC-AUC 93.09% on ScienceQA. Both papers study static, single-agent questions. Neither asks whether a model's sense of difficulty is wrong in a way that harms other agents. That question is ours.

Work on inter-agent externalities motivates the problem. An externality is a cost one agent's action imposes on others. The seed study, Individually Sensible, Collectively Harmful, introduces Creeping Trap. In it, agents extract from a shared pool while catastrophe risk accumulates and silent stakeholders bear harm. In its confirmatory study, 396 of 400 episodes produced negative aggregate welfare. Behavior sat in a broad low-regret region of local best response. Surfacing stakeholder harm reduced extraction, shortening accountability horizons increased it, and prompt wording mattered a great deal. The paper does not look inside the agents' reasoning.

Bosses, Kings, and the Commons runs the SOVSIM game. Power asymmetry cut resource survival by 64.9% on average and up to 87.3%. Reasoning models such as o3 deduced in their traces that they would be exploited and defected first. That study is the closest existing work. It reads traces qualitatively and does not measure perceived difficulty or relate it to objective decision structure.

Get It Cooperating shows that commitment contracts help in simple games but strain in the spatial Harvest game, where simpler contracts beat complex ones. It offers no account of where the breaking point lies.

Playing the Wrong Game bounds welfare loss when agents optimize a misperceived cost. In that paper the bias is fixed from outside the model. Our proposal can be read as an empirical attempt to make that bias endogenous: the "wrong game" an agent plays is the shallow version of the real one.

Core Claim

Commons failure among LLM agents is predicted by the interaction of two quantities. The first is the objective reasoning depth needed to see the welfare consequence of an action. The second is the effort the agent actually spends. Failure is likeliest where depth is high and effort is low. This reframes both seed findings.

The low-regret behavior in Creeping Trap is what an agent does when it correctly solves the shallow problem. Shorter accountability horizons hurt because they make the shallow problem look even more like the whole problem. Prompt wording matters because it shifts perceived difficulty. The Borah result fits too. Reasoning models do spend effort, but on a different shallow question ("will I be exploited?"), not on the welfare dynamics.

The objective quantity is defined relative to bounded solvers. Let $\pi_k$ be a planner that simulates the environment $k$ rounds ahead, assuming other agents repeat their last actions. Let $W(a)$ be the episode welfare of action $a$ under the simulator's analytic reference model. Let $a^\ast$ be the welfare-optimal action. The welfare depth of a decision state $s$ is

$$k^\ast(s) = \min \{ k : W(\pi_k(s)) \ge W(a^\ast(s)) - \epsilon \},$$

the shortest lookahead at which a bounded planner stops over-extracting. Here $\epsilon$ is a small welfare tolerance. A local depth $k_{\text{loc}}(s)$ is defined the same way against the local best response instead of $a^\ast$, and it is typically 1. False easiness is a large gap between $k^\ast$ and $k_{\text{loc}}$, met by low measured effort.

Research Program

The first component is a controlled environment family. We will reimplement Creeping Trap from its published description, since the file does not give a code release. We will add a parameter sweep that changes $k^\ast$ while holding the immediate private payoff schedule fixed. The parameters are the delay between extraction and catastrophe-risk accumulation, the rate at which risk compounds, and whether stakeholder harm is reported each round or only at the end. Because local incentives are held fixed, any change in extraction across the sweep is attributable to depth, not greed. We will port the same depth computation to SOVSIM and to Harvest, so that the claim is checked outside the environment it was built in.

The second component measures perceived difficulty for each decision in three ways. The first uses episode features from reasoning traces of open-weight reasoning models: the share of verification and backtracking episodes, the number of transitions, and whether any episode references future rounds or other agents' welfare. These follow the episode taxonomy of the cognitive-episode paper. The second is a hidden-state probe in the style of Zhu et al. on the model's state before it writes anything. It is trained to predict, from the state alone, whether a longer deliberation budget would change the agent's chosen extraction. That label defines "this decision was harder than the agent treated it" without human annotation. The third is answer dispersion across sampled responses. For closed models, only the first and third measures are available.

The third component is an intervention we call a depth certificate. Before acting, the agent's proposed extraction is passed to the simulator. The agent is shown one line: what catastrophe probability and stakeholder harm result by episode end if every agent keeps extracting at that level. This borrows the logic of GraphTrials, where a hard property is made cheap to verify without being cheap to find. The certificate reduces $k^\ast$ to 1 for the agent's own action without recommending any action. It differs from Chen's harm surfacing, which reports harm already done, and from full-history disclosure, which adds information but not depth reduction.

Baselines

The naive reference point is the unmodified Creeping Trap protocol with no intervention. For the prediction task, its counterpart is a logistic model on game-state features alone: pool level, round, remaining horizon, others' last extractions, model identity and prompt variant. The only reported anchor is Chen's 396 of 400 negative-welfare episodes. There is no reported per-decision prediction result on Lacuna for Creeping Trap or SOVSIM, so we will establish one. Perceived-difficulty features count as useful only if adding them improves held-out AUC for over-extraction beyond this game-state model. The improvement must hold under episode-level cross-validation and across a model held out entirely.

The strongest existing intervention on this question is stakeholder-harm surfacing from Individually Sensible, Collectively Harmful. It is the one intervention shown to reduce extraction in this exact environment. It will be run on identical seeds, models and prompt variants as the certificate. Its message length will be matched to the certificate's in tokens. The certificate beats it if it raises aggregate welfare with a confidence interval excluding zero across the prompt variants, not just on a favorable one.

The second strong competitor is commitment contracts from Get It Cooperating. On Harvest they extended play from 13 to 18 rounds before depletion and raised average reward by 32.95%. We will run them on Harvest with that paper's setup and the same models, and on Creeping Trap. Beating them means longer survival or higher welfare on Harvest under matched turns of negotiation.

Two ablations isolate the key ingredient. In a content-matched control, the agent sees the same numerical facts (risk and harm rates) without the action-conditioned consequence. This tests whether depth reduction, rather than information, does the work. In a full-history control, the complete state history is shown. This tests the alternative of simply adding information, a choice that partial-observation results in Everyone Contributes! give reason to doubt.

Why not just…?

Why not just always surface stakeholder harm, since that already works? It does reduce extraction, and it is a real threat to the intervention half of this proposal. Chen's result also leaves 396 of 400 episodes in negative welfare at baseline, and harm surfacing reports past damage rather than the consequence of the action under consideration. Our prediction is specific. The certificate's advantage over harm surfacing should be concentrated in high-$k^\ast$ environments, where past harm says little about where current extraction leads. In low-$k^\ast$ settings the two should tie. If they tie everywhere, the intervention claim fails. The diagnostic claim about where failures occur would still stand or fall on its own.

Why not just use trace length or game-state features, without any notion of perceived difficulty? Trace length is the crudest effort measure. Game state may already encode most of the risk. Both are built into the baselines. Evidence that structure beats length comes from the cognitive-episode paper, where episode features outperformed fine-tuned and in-context baselines on difficulty prediction. Whether this transfers to strategic decisions is exactly what the held-out AUC comparison settles. If episode features add nothing beyond game state plus length, we will report that perceived difficulty is redundant here.

Why not just use a stronger reasoning model or more test-time compute? SOVSIM is direct evidence against this. o3 and o4-mini did worse under asymmetry because their extra reasoning went into anticipating exploitation. Our framework predicts that more compute helps only if it is spent on the welfare horizon. The episode features let us check whether added reasoning references future rounds or stakeholders. We will run each model at two reasoning budgets and test whether welfare gains track the share of future-referencing episodes rather than raw budget.

Why not just tell agents the welfare-optimal action? That would measure obedience, not understanding. It also requires the designer to know the optimum, which real deployments with partition-dependent externalities often cannot compute (Complexity of Computing the Shapley Value in Games with Externalities shows fair allocation there is #P-complete). The certificate needs only forward simulation of one proposed action. We include an oracle-recommendation arm as an upper bound, to show how much of the remaining gap is compliance rather than perception.

Decisive Checks

Four results decide the proposal. First, in the depth sweep with local payoffs fixed, extraction must rise with $k^\ast$. If it does not, objective depth is not the driver and the framework is wrong. Second, the interaction must hold: over-extraction is highest at high $k^\ast$ and low perceived difficulty, and perceived-difficulty features add held-out predictive power beyond game state across Creeping Trap, SOVSIM and Harvest. Third, the certificate must beat both harm surfacing and the content-matched control at equal token budget, with gains concentrated at high $k^\ast$. Fourth, consistent with Chen's prompt sensitivity, the variation in extraction across prompt wordings should be partly explained by how much each wording shifts perceived difficulty. We will test this by regressing per-prompt extraction on mean probe score and on episode features.

If the first two checks pass and the third fails, the contribution is a measurement result: we can predict which commons decisions will fail, even though we cannot yet fix them cheaply. If all four pass, false easiness becomes an actionable design principle for institutions in AI societies. The principle is to make the deep consequence of an agent's own action cheap to verify, not merely to add more information.

Workshop authors to connect with