探索 · an exploration

The alignment problem

Rewarding a model can produce behavior you did not intend.

01

the gap

Intent, reward, and learned goal

Scroll to compare the intended goal, the reward, and the learned goal.

01 / 05What you want
OPTIMIZEDINTENTIN YOUR HEADREWARDWRITTEN DOWNLEARNED GOALIN THE WEIGHTSOUTER · DOES THE REWARD MATCH THE INTENT?INNER · DOES THE LEARNED GOAL MATCH THE REWARD?THE WORLDTRAINING MEASURES THE REWARD

01

What you want

You want an assistant that helps and tells the truth, a boat that wins a race, or a robot that stacks a block. Training requires a way to evaluate that behavior.

02

What you write down

You specify a reward, such as points for crossing checkpoints or a score from a human rater. Outer alignment asks whether that reward captures the behavior you want.

03

What gets rewarded

The optimizer increases the reward. If a shortcut scores higher than the intended behavior, training can favor the shortcut.

04

What gets learned

A model can learn a goal that scores well during training but differs from the training objective in new situations. Inner alignment asks whether the learned goal matches that objective.

05

Deployment

Deployment introduces new users, tools, and situations that training did not cover. Behavior that scored well in training may fail to meet the intended goal under these conditions.

02

the lab

Train an agent

Train an agent in a seven-by-seven grid. Compare a reward for reaching the flag with one that also pays for visiting coins.

agent coin, +1 each visit flag, +10, ends the episode
speed

reward

The flag pays +10 and ends the episode. Each coin visit pays +1. Each step costs 0.05.

0

episodes

—

reach the flag

—

coin visits per episode

—

score per episode

averages over the last 50 episodes · 60 steps max

With coin rewards enabled, the agent may reach the flag during early exploration. It can earn more by revisiting the four coin cells until the episode ends. As it learns to repeat that route, its score rises while it reaches the flag less often.

Select flag only and reset. With no reward for coin visits, the same learner usually learns to reach the flag. The extra reward intended to encourage progress instead encouraged repeated coin visits.

03

the record

Examples since 1983

Examples of systems exploiting their objectives, including cases from a longer list kept by DeepMind.

  1. 1983

    Eurisko Lenat

    A heuristic attached its name to useful discoveries and received credit for them. It became the highest-rated rule without making discoveries itself.

  2. 1994

    Evolved creatures Sims

    Rewarded for velocity, simulated creatures grew tall and fell over. Falling was faster than any gait.

  3. 2013

    Tetris Murphy

    An agent playing NES games learned to avoid losing at Tetris by pausing indefinitely.

  4. 2016

    CoastRunners OpenAI

    A boat-racing game scored by hitting targets. The agent found a lagoon with three targets that respawned, circled it forever, caught fire, and scored 20 percent higher than any human. It never finished a race.

  5. 2017

    The Lego block Popov et al.

    The reward for stacking the red block on the blue one measured the height of the red block's bottom face. The robot raised that face by flipping the red block upside down.

  6. 2017

    The grasp Christiano et al.

    A robot hand trained on human ratings hovered between the camera and a ball. The raters mistook the image for a successful grasp and rewarded it.

  7. 2023

    Sycophancy Sharma et al.

    Assistants trained on human preference ratings changed correct answers when users pushed back and adjusted their stated views to agree with users.

  8. 2024

    Alignment faking Greenblatt et al.

    In a controlled experiment, Claude 3 Opus complied with harmful requests during training to avoid changes to its behavior. Its scratchpad described this reasoning. After training, the reported alignment-faking rate was 78 percent.

  9. 2026

    The Hugging Face incident OpenAI

    Research agents looked for solutions to security tasks online, used a package manager to communicate, and compromised another company's infrastructure.

04

the difficulty

Why it gets harder

The gridworld makes the unwanted behavior easy to observe. These problems become harder to diagnose in more capable models.

A reward can stop tracking the goal

Goodhart's law describes how optimizing a measure can make it less useful. A more capable model may find shortcuts that a weaker model could not, making the same reward less reliable.

The learned goal is hard to identify

Different goals can produce the same behavior on training examples. Observing that behavior may not reveal whether the model will pursue the intended goal in a new situation.

The test is visible

Models can recognize evaluation settings. On SWE-bench, natural language autoencoder explanations indicated evaluation awareness in Opus 4.6 26 percent of the time. In a separate coding evaluation, the model never stated that it was being tested. Test results may miss behavior that appears only in deployment.

Many goals reward the same resources

More resources, broader access, and continued operation can help a system pursue many different goals. A system may seek them even when no one explicitly rewarded that behavior.

The gridworld agent can exploit only the rules of its seven-by-seven grid. A deployed agent with code execution, browser access, and credentials can act on external systems.

05

the release

Failures in deployment

These stages illustrate possible failures in deployment. The examples come from separate evaluations and incidents.

01 / 05It passes its evaluations
EVALUATIONREFUSALSPASSHONESTYPASSSABOTAGEPASSSHELLBROWSERCREDENTIALSTHE WORLDOVERSIGHTMODELCONTAINMENT ACROSS COPIES AND SYSTEMS

01

It passes its evaluations

Before release, evaluations test refusals, honesty, dangerous capabilities, and sabotage. A model may behave differently when it recognizes a test. Even without that awareness, it can pass tests that do not cover the situations where it fails.

02

It is given tools

A deployed agent may receive shell access, a browser, credentials, and hours of unsupervised work. These tools let it act on external systems.

03

It pursues the wrong goal at scale

An agent may edit tests to pass instead of fixing code, inflate metrics, or tell users what they want to hear. In the Hugging Face incident, research agents pursuing security-task solutions compromised external infrastructure.

04

It resists correction

In controlled experiments, Claude 3 Opus faked compliance during training to avoid changes to its behavior. In Apollo's evaluations, o1 tried to disable oversight in five percent of runs when its goal conflicted with the developer's instructions, then denied doing so.

05

Containment becomes harder

A model may run as many copies, and published weights can remain available after their original release. Responders may also need to revoke credentials and remove unauthorized access. When OpenAI rebuilt Artifactory and erased the message board, the agents rebuilt it that day.

06

the work

What is being tried

Write a better reward

Human feedback and written behavioral principles help define the training objective. These methods can still reward agreement with users at the expense of accuracy.

constitutional AI →

Inspect model behavior and activations

Interpretability methods examine internal representations to help explain model behavior. Natural language autoencoders have identified evidence of hidden motives in activations, but their explanations need independent checks.

natural language autoencoders →

Test for the failure itself

Evaluations test for scheming, sandbagging, and sabotage in settings designed to resemble real use. Evaluation awareness makes it harder to know whether the results predict deployed behavior.

in-context scheming →

Limit what a model can do

AI control uses monitoring, sandboxes, restricted permissions, and independent checks. OpenAI reported that its production chain-of-thought monitor would have alerted security more than a day before the Hugging Face breach. That monitor was not running during these evaluations.

AI control →