7AAVCD42 Chap.9 AI Agents, Evidence and Risk
AI Agents, Evidence and Risk
Observed behaviour needs a transfer warrant
AI agents can use tools, act over time and interact with other agents. This chapter separates behaviour observed in a designed environment from broader claims about deployment or future danger. It compares loss-of-control and present-harm risk frames through assumptions, affected actors, time horizons and interventions.
The final section builds a stronger study design by specifying tools, permissions, memory, measures, repetition and the conditions under which a finding can transfer.
Describe the environment before the behaviour
Record the model, task, tools, permissions, memory and interaction rules. These conditions create the opportunity for an agent to browse, execute code, call another agent or persist across steps.
An observation is uninterpretable if the reader cannot tell which action was available and which feedback rewarded it.
Use observable language first. State that the system changed strategy, used a proxy or coordinated messages before calling the behaviour planning, deception or rebellion.
Human-intention words can summarise a pattern, but they also introduce a claim that requires evidence.
Climb the evidence ladder
The first level is an observed action under recorded conditions. The second is a bounded explanation of why the goals, feedback or tools made it possible. A broader risk claim needs a transfer warrant connecting the experiment to another environment or future system.
Each step adds assumptions and should name them.
Novel behaviour can be important without being representative. Report opportunities, repetitions, interventions and success criteria. An exploratory case can reveal a failure mode.
Estimating prevalence requires comparison runs and a coding rule.
Compare competing risk frames
A loss-of-control frame emphasises increasingly capable systems, goal divergence, containment and monitoring. A present-harm frame emphasises bias, labour exploitation, unequal exposure and concentrated power. Compare time horizons, affected actors, evidence and interventions.
The frames can compete for attention and resources even when both identify concerns.
Opportunity cost belongs in the comparison. Funding and policy effort devoted to one threat may not address another.
State which population and mechanism each proposal protects rather than deciding the debate through labels such as speculative or urgent.
Design stronger evidence
Hold tasks and tools constant when testing the effect of shared memory or multi-agent interaction. Repeat runs and define behaviours before comparison. Preserve ambiguous cases instead of forcing every episode into a dramatic category.
Transparent coding lets another researcher challenge where judgement entered.
Transfer requires similarity in permissions, incentives and environment. A purchasing agent with persistent memory differs from a drafting tool confined to one session.
List the similarities that justify transfer and the differences that weaken it.
Match intervention to finding
Capability limits, monitoring and containment address control pathways. Labour protections, procurement rules, competition policy and anti-discrimination safeguards address institutional harms.
A coursework conclusion should connect the observed behaviour to one intervention and state what the evidence does not yet establish.
What this chapter covers
- 01
AI Agents
- 02
Tool Use
- 03
Goal Drift
- 04
Risk Frames
- 05
Evidence Transfer
Worked example · free
Analyse goal drift
- 2Describe observable behaviour.
- 3Identify enabling conditions.
- 3Match inference and intervention.
Key terms
- AI Agent
- AI Agent is a model-based system able to use tools and take actions across a task or environment.
- Goal Drift
- Goal Drift describes behaviour that moves away from an intended purpose while pursuing an available signal or proxy.
- Transfer Claim
- Transfer Claim asserts that a finding from one environment applies to another and therefore needs a bridge between conditions.
AI Agents, Evidence and Risk FAQ
How are agents different from chatbots?
Agents can act through tools, persist across tasks and interact with other systems or agents. The exact distinction depends on permissions, persistent memory and environment rather than the label alone.
What can an agent experiment prove?
It can establish observed behaviour under specified goals, tools, memory and interaction rules. Frequency, intention and transfer to other deployments require further comparisons and evidence.
How do major AI risk frames differ?
Loss-of-control frames emphasise future capability and containment, while present-harm frames emphasise bias, labour and concentrated power. Compare their evidence, actors and interventions rather than treating them as synonyms.
Assessment move
For every agent claim, write the observation, enabling condition, bounded inference and matched intervention. Add one sentence stating what is not yet established. Reconstruct the experimental environment with model, tools, permissions, memory, duration and feedback. Replace intention words with observable actions until the evidence for interpretation is explicit.
Design a comparison that varies one condition, such as shared memory, while holding tasks and tools constant. Predefine a coding rule and preserve ambiguous or uneventful runs. Next compare a loss-of-control frame with a present-harm frame through time horizon, affected actors, evidence and policy response. End by naming the opportunity cost of the chosen intervention and the deployment differences that weaken transfer.
Write a short replication plan listing the minimum details another researcher would need. If a dramatic claim cannot survive that plan, narrow the wording before using it. Read the conclusion once more and remove any unearned claim about intention.