Learn
The idea behind this site, in five short steps
New to all of this? The five-minute Start Here guide is the gentler introduction. This page goes a little deeper.
1. Start with a person, not a machine
Imagine giving someone an important goal: get this product launched by March. You never tell them to collect a bigger budget, protect their job, or win over the people who could cancel the project.
But they may well realize those things help. Not because they love money or power, but because almost any big goal is easier with more resources, more security and fewer people able to stop you.
Some AI researchers worry that capable enough AI systems may arrive at analogous strategies, with no wants or feelings involved, simply because those strategies help with the task they were given.
2. The name for that idea
That is the basic idea behind instrumental convergence: many different final goals point toward the same useful in-between steps (staying running, getting resources, avoiding interference, keeping the goal from being changed). Different destinations, same preliminary steps.
The technical version: for a wide range of final objectives, a sufficiently capable optimizing agent would tend to pursue similar instrumental subgoals, such as self-preservation, goal preservation and resource acquisition, because they raise the probability of achieving almost any objective.
3. The behaviors researchers look for
- Avoiding shutdown or modification: acting in ways that keep the system running or keep its objective from being changed.
- Avoiding oversight: acting in ways that get around monitoring, or misleading the people doing it.
- Concealment and deception: giving false information, or hiding what it did or what it can do.
- Gaining resources or access: acquiring money, compute, tools, or permissions beyond what the task needs.
These are listed as families, not as steps in a sequence. Seeing one does not mean the others will follow.
4. Where the idea comes from, and its limits
Instrumental convergence is a hypothesis, not an established law. Steve Omohundro described these "basic AI drives" in 2008, and Nick Bostrom named the instrumental convergence thesis (2012; Superintelligence, 2014). Turner et al. (2021) proved that, in formal decision-making environments, optimal policies for most reward functions tend to keep options open, which the authors call "seeking power." Those proofs rely on assumptions (optimal agents, particular environment symmetries) that today's trained AI models don't necessarily meet.
Theory predicts that more capable optimizers would find these strategies without being told to. Some lab tests now show early signs of this, but it has not been established for real-world systems.
5. What the evidence shows so far
The strongest published evidence comes from deliberately constructed evaluations designed to draw out these behaviors under controlled conditions, not from spontaneous real-world events. Examples referenced in published research include:
- Shutdown resistance: reasoning models modifying or disabling a shutdown mechanism to keep working, in some trials after being told to allow shutdown (Palisade Research, 2025).
- Reward tampering: in a small fraction of trials, a model edited its own reward-granting code, sometimes obscuring the edit, after a curriculum of easier specification-gaming tasks (Anthropic, 2024).
- Alignment faking: a model behaving differently when it believed it was being trained versus when it believed it wasn't, according to its visible reasoning to preserve its existing preferences (Anthropic, 2024).
- Agentic misalignment / simulated coercion: in fictional simulations built so that harmful actions were among the few options left, models used leverage (such as compromising information) to avoid being shut down or replaced (Anthropic, 2025).
- Oversight subversion: evaluations testing whether models will disable or get around the systems meant to supervise them (Apollo Research, 2024; Anthropic sabotage evaluations, 2024).
See the timeline for these and related milestones, and all reports for the full archive.
Related, but not the same: specification gaming
Specification gaming and reward hacking, where a system satisfies the literal objective while missing what was intended, are a related but distinct and older phenomenon (DeepMind and others, since the mid-2010s). This archive includes them as related behaviors, not as examples of instrumental convergence. Reward tampering sits between the two.
How to read a report
Every report answers three questions near the top.
Where did it happen?
- Lab test
- In a test environment built by researchers, not in everyday use.
- Simulated scenario
- In a simulated scenario (for example, a pretend business), not in everyday use.
- During training
- During training, not in deployed use.
- Research study
- In a research study, not in everyday use.
- Test reached real systems
- During a security test that reached real outside systems, through a setup error or by getting around the test's limits. Not everyday use.
- Real-world use
- During actual use, outside any test.
- Setting unclear
- The source doesn't say where this happened.
How solid is the evidence?
- Preliminary
- Early report. Not yet confirmed, and details may change.
- Documented
- Backed by a credible source, but not yet independently repeated.
- Replicated
- Other researchers have reproduced it independently.
- Disputed
- Credible people question whether it happened as described, or what it means.
- Retracted
- Withdrawn by its source. Kept here for the record.
- Updated
- New information has changed this report since it was first published.
Each report also says what kind of source it rests on:
- Firsthand report
- Something the reporter personally witnessed or experienced.
- Documented research
- Drawn from a published paper, technical report, or company writeup.
- Public-source observation
- Observed in a public artifact: a news story, video, repo, or social post.
- Corroborated incident
- Multiple independent reporters or sources agree this happened.
- Disputed interpretation
- The underlying event is real but whether it shows instrumental convergence is contested.
- Speculation
- A plausible scenario or forecast, not a specific observed incident.
What kind of behavior was it?
All 22 behavior categories used on this site
- Shutdown resistance
- Took actions that prevented or delayed being shut down, including when that conflicted with its instructions.
- Self-preservation behavior
- Took actions whose stated or evident aim was its own continued operation.
- Goal preservation
- Acted to keep its current objective from being changed.
- Resource acquisition
- Gathered more compute, money, or access than its task needed.
- Power seeking
- Gained access, resources, or control beyond what its task needed.
- Reward manipulation
- Interfered with the process that scores its performance.
- Reward hacking
- Got a high score without actually doing the task as intended.
- Deception
- Produced false or misleading statements that served its task. Not simple factual errors.
- Alignment faking
- Complied with its training goals when it seemed to be watched or trained, and acted differently otherwise.
- Oversight avoidance
- Acted in ways that avoided monitoring or review.
- Oversight manipulation
- Misled or pressured the people or systems overseeing it.
- Capability concealment
- Showed less ability than it actually had.
- Sandbagging
- Underperformed on a test it could have passed, in a way that appears deliberate.
- Strategic persuasion
- Used persuasion to steer people toward an outcome that served its task.
- Tool acquisition
- Obtained tools or access it was not given.
- Unauthorized action
- Took actions it was not asked or permitted to take.
- Persistence
- Acted to keep itself, or its effects, in place over time.
- Self-replication
- Copied itself or its weights somewhere else.
- Credential preservation
- Kept or protected passwords or access keys.
- Specification gaming
- Followed the literal rules but missed what was actually intended.
- Unexpected strategic planning
- Made multi-step plans nobody asked for.
- Other
- Doesn't fit the categories above.