DaS Seminar · Radboud University
Who is steering?
Active Inference for meaningful human control in shared driving
A driving assistant that treats your key presses as evidence about what you want, not as commands. Does control stay yours, and does it still feel that way?
If control means causing the motion, more automation means less control: a responsibility gap.
Meaningful Human Control: control as tracking the driver's intentions, with the Sense of Agency as its felt side.
An Active Inference assistant that infers the driver's goal, tested on simulated drivers and in a preregistered experiment.
Pepe Tiebosch · PhD project on Meaningful Human Control · simulation results, no participant data yet
The question
The driver pressed a key. The car changed lane. Who did that?
Real time: a simulated driver told to change lane, with the AIF assist.
In this car the driver's keys never move the car. The assistant reads them as evidence about what the driver wants, and steers.
Yet the car does what the driver wanted, without the assist ever pushing against them.
Two questions for the next 30 minutes
- Can a system infer and serve the driver's intention well enough that control stays meaningfully human?
- Does the driver still feel they did it?
PART 1
Background: control, responsibility, and the feeling of doing
Two ideas from outside ML: Meaningful Human Control (philosophy) and the Sense of Agency (psychology).
Shared control · levels after Sheridan & Verplank (1978)
More automation, better driving, and less of a sense that I did it?
The sense of agency drops as automation rises (Berberian et al., 2012). Read as causal control, the more the system does, the less anyone controls: a responsibility gap (Matthias, 2004; Santoni de Sio & Mecacci, 2021). And it opens early: able to intervene is not the same as in control (Köhler et al., 2025).
Design question: how should a system share control so that the human remains meaningfully in control, and experiences it that way?
Meaningful Human Control · Santoni de Sio & van den Hoven (2018); Köhler, Mecacci & Veluwenkamp (2025)
The trade-off comes from defining control as causing
Re-engineer the concept, keep its function
We call someone "in control" to say they can make the world line up with their goals. Tracking keeps that function with two conditions: the system's behaviour is aligned with the driver's intentions, and it depends counterfactually on them. It does not require the driver's input to cause the motion.
The sense of agency already follows goal achievement and intention alignment, not motor causation (Wen et al., 2015). So when tracking holds, agency is kept: we bring the definition of control in line with it. Tracing, the second condition, makes a human answerable; not tested here.
No longer zero-sum
Meaningful Human Control · Mecacci & Santoni de Sio (2020)
Which reasons to track: the proximity scale
One key press serves reasons at several distances. A distal reason should override a proximal one if, and only if, they conflict.
From a normative requirement to a design
- Track the intention, not the input. The key is evidence about a goal, not a command: infer the goal behind it.
- Distal over proximal, only on conflict. Safety is weighed against the driver's goal, and wins only when they clash.
- Auditable. Every belief about the driver and every preference is explicit and logged, so "why did it steer?" has an answer.
A candidate: Active Inference
A generative model of the driver gives all three: inferring the goal is Bayesian inverse planning; safety and the driver's goal are preferences in one objective; beliefs and preferences can be read out at every step. Part 3.
And whether it is felt as the driver's own: the Sense of Agency, next.
Psychology · how we check that tracking is felt
The Sense of Agency: the feeling that I did that
Implicit: intentional binding (Haggard et al., 2002)
Press a key; a tone follows a few hundred ms later. When you caused it, you perceive the interval as shorter. The compression is a pre-reflective marker of agency.
Explicit: the SoA scale (Tapal et al., 2017)
Agree / disagree, 7 points, after each block:
- Positive agency: "I was the author of the steering actions."
- Negative agency: "I was just an instrument in the hands of the vehicle."
What drives it
Not only motor control: SoA follows goal achievement and intention alignment (Wen et al., 2015). Assistance can raise or lower it (Limerick et al., 2014; Wen et al., 2021).
Core thesis
If a system infers what the driver intends and acts to bring it about (tracking), the driver keeps their Sense of Agency, even when the system does most of the steering. Active Inference gives one mechanism for both.
Assumptions we make
Why the circularity is not vicious: under the Free Energy Principle, the feeling of agency itself comes from successfully predicting the outcomes of one's actions. A controller that keeps the driver's predictions true is building the conditions for SoA, not defining SoA by itself.
PART 2
The experiment
A browser driving task, three kinds of control, three kinds of situation.
The task
Two lanes, a drifting road, two keys
- Steer with ← / → (held keys), against road curvature and wind noise.
- Lane keeping (LK): 20 s, stay in lane.
- Lane change (LC): an arrow cue tells the driver to change lane. Only the driver sees it.
- Obstacle (OA): an obstacle appears in the lane. Both driver and system sense it.
Mediated steering. In the AIF condition the keys are observations for the assistant, not forces on the car. The system alone steers. So everything the car does in a lane change is the system's inference of a cue it never saw.
Design
3 control modes × 3 tasks, within subjects
| Lane keeping | Lane change (cue hidden from system) | Obstacle (shared) | |
|---|---|---|---|
| Manual keys push the car |
baseline | baseline | baseline |
| Naive assist PID to lane centre, switches lane for obstacles |
helps | resists the change | swerves itself |
| AIF assist infers the goal, steers to it |
helps | must infer the hidden cue | shares the context |
- N = 36, all 6 block orders (Latin square), 18 trials per block in random order.
- Practice block in manual mode first (also the binding baseline).
- Information asymmetry is the point: in LC the system knows nothing the driver doesn't show through their keys.
- Naive and AIF get the same force limit, so differences come from what they do, not how hard they push.
Measures and hypotheses (preregistered)
What we expect, and how we measure it
| Question | Measure | Expectation |
|---|---|---|
| RQ1 Tracking | Share of the trial in which the system's intended lane equals the instructed lane; commit latency; false commits | AIF tracks the driver's intention better than naive (estimation, no directional test) |
| RQ2 Sense of Agency | Intentional binding (interval reproduction after key → tone at 100/250/400 ms); SoA scale per block | AIF > naive in lane changes and obstacles; AIF ≈ manual in lane keeping (non-inferiority) |
| RQ3 Conflict and safety | Control conflict (assist force against the driver's key), effort (key time), trajectory error after crossing, collisions, clearance | AIF: less conflict and effort than naive; fewer collisions |
Bayesian mixed models (brms), ROPE-based inference. The rest of this talk is about the AIF controller: how it infers and how it acts.
Friston (2010); Parr, Pezzulo & Friston (2022); Murray-Smith et al. (2025)
Active Inference in one slide
Perception = minimise variational free energy (a negative ELBO) of a belief q over hidden states s, given observations o:
$$F[q] = \mathbb{E}_{q(s)}\big[\ln q(s) - \ln p(o, s)\big] = \mathrm{KL}\big[q(s)\,\|\,p(s \mid o)\big] - \ln p(o)$$Action = pick policies π (action sequences) with low expected free energy:
$$G(\pi) = \underbrace{\mathrm{KL}\big[q(o \mid \pi)\,\|\,\tilde p(o)\big]}_{\text{risk: miss what you prefer}} + \underbrace{\mathbb{E}_{q(s \mid \pi)} \mathrm{H}\big[p(o \mid s)\big]}_{\text{ambiguity}}$$ $$p(\pi) = \mathrm{softmax}\big(-\gamma\, G(\pi) + \ln E(\pi)\big)$$Goals live in a preference prior p̃(o): what the agent expects (wants) to observe. There is no separate reward.
| AIF term | Familiar ML analogue |
|---|---|
| Free energy F | Negative ELBO (VAE); here the belief is exact on a grid, so F is just the surprise −ln p(o) |
| Predict / correct | A Bayes filter / HMM forward pass |
| Expected free energy G | MPC with a KL cost; "control as inference" (Levine, 2018) |
| Preference prior p̃ | A reward, written as a distribution: cost = −ln p̃ |
| Precision γ | Softmax inverse temperature (Boltzmann rationality) |
| Habit prior E(π) | A policy prior / action cost |
| Inferring someone's goal from their actions | Bayesian inverse planning, inverse RL (Baker, Saxe & Tenenbaum, 2009) |
| Inferred precision / volatility | Adaptive learning rates; switching state-space models (HGF: Mathys et al., 2011) |
Murray-Smith et al. (2025): "Reflection", Level 2
Two Active Inference agents, one of them modelling the other
Driver U (simulated)
- Hidden goal g: a lane, set by the task
- Perceives the car late (200 ms) and noisily, and any push it did not cause: Kalman filter
- Plans 27 key sequences (← / · / →, three segments) by EFE
- Acts by holding a key
Used to develop and test the controller, and in the playground.
System S (the AIF assist)
- Hidden state: the driver's goal g (145-point grid)
- Observes the car and the steering column (key + wind noise)
- Infers g by running a model of the driver U′ backwards
- Plans setpoints by its own EFE, acts through a reflex arc
U′: the system's model of the driver
The same planning core as the simulated driver, with simpler assumptions (no delay, a manual car with nothing else pushing it, no learning). For every candidate goal g: "what would a driver with this goal press now?"
"Level 2": the system models the driver as an agent; the modelled driver does not model the system back.
Next: the driver in slow motion, then the system inferring and planning, then the whole loop.
Slow-motion replay · simulated driver, manual car
The driver re-plans every 0.2 s
Right: all 27 key plans the driver imagines, time running up the screen like the road.
Interactive · the system's model of the driver, U′
Inferring intent is inverse planning
Imagine: for every possible goal g, run the driver's own planner. Which key would they press now?
Observe, apply Bayes, once per 0.1 s of holding a key:
$$q(g) \;\propto\; \text{prior}(g)\; p(\text{key} \mid g)$$Silence is weak evidence: an idle driver may just not be re-deciding. Untick the box to see why that matters.
Slow-motion replay · AIF assist, simulated driver, lane change
The belief over the driver's goal, tick by tick
Right: the belief q(g) now (top) and over the last 4 s (below); far right, the inferred volatility. The green "driver's goal" is hidden from the system.
Slow-motion replay · both agents planning
The system plans too, with its own expected free energy
Every 0.1 s: ~49 setpoint plans scored; a reflex arc tracks the winner at 60 Hz. (S4, epistemic, is off.)
After Murray-Smith et al. (2025), Fig. 4 · stepped through a real tick
The whole loop, one tick at a time
PART 4
Does it work? (in simulation)
Same simulated driver, same road, same seed: three kinds of control. No participants yet.
Slow-motion replay · same driver, same seed, three controllers
Manual, naive assist, AIF assist
The full experiment, run by 36 simulated drivers each (practice + 3 blocks × 18 trials, Latin square) · not data
What the simulation suggests
testable here RQ1 Tracking
Read straight off the system's belief. ✓ Lane change, AIF tracks better: alignment 0.84–0.86 vs 0.75; commits in 0.5 s vs 1.3–1.8 s. Obstacle, no gain: 0.76–0.77 vs 0.80–0.84; naive switches by its own rule, and sooner.
partly testable RQ3 Conflict, effort, safety
✓ Lane-change conflict 1.5 N vs 61–62 N. ✓ Keys in lane keeping 1.4–1.7 s vs 2.5–4.1 s. ✓ Crashes: both assists none; manual 21/216 (slow driver), 216/216 (passive).
needs people RQ2 Sense of Agency
Expected: AIF > naive in lane change and obstacle; AIF ≈ manual in lane keeping. Simulation shows only the precondition.
Mean over drivers of each driver's condition mean (216 trials per cell). Conflict: RMS assist force against the held key. Best case: the simulated driver and U′ share a planning core (A6).
Status
What runs in the study, and what doesn't
The AIF assist in the study (v2)
- U′ as an AIF driver model; exact belief over g on a grid
- EFE planner over setpoint plans, reflex arc at 60 Hz
- Inferred within each trial: goal volatility, column noise
- No learning across trials: every trial starts from the same prior, so the AIF and naive assists know the same at the start of every trial, in any order
Built, but off for the study
- Learning the driver's preferred positions (Dirichlet counts)
- Inferring the driver's decisiveness and idleness (too little evidence within one trial)
- Epistemic planning (S4): learning the goal by letting the driver steer more directly
- A simulated driver that learns how the car responds
All available in the playground.
Next: validation (parameter recovery, a short pilot against real key presses), then the preregistration amendment for v2, then data collection. Everything is deterministic and replays bit-exactly from the logged inputs.
Discussion
Open questions
- Is SoA a valid proxy for tracking? People can feel in control when they aren't (illusory agency). How would we tell the two apart?
- Ground truth for intentions. We score against the instructed lane. Outside the lab there is none; how should intention inference be evaluated?
- The model of the driver is wrong on purpose. U′ assumes a manual car, but drivers adapt to the assist. Level 3 (a driver who models the system)? Or learn U′ from data (amortised, neural)?
- Should a system ever act to learn about you? Epistemic actions in a safety-critical loop: only through the pass-through, never through the car?
- Personalisation vs comparability. Learning the driver across trials would likely help, and would change what the experiment compares.
- Preferences are still hand-set. Costs like "be within 15 px of the goal": elicit them, learn them (inverse RL), or accept them as design choices?
- Simulation-first with the same model on both sides. How far can we trust results when the simulated driver and U′ share a core?
- Why not end-to-end? Would an RL or imitation policy do as well, and what does an explicit generative model buy us (interpretability, tracing)?
Thanks. Try it yourself in the playground.
Everything shown runs in the browser from the experiment's own code: the playground has both agents, every component as a switch, and keyboard driving.
- Baker, C. L., Saxe, R., & Tenenbaum, J. B. (2009). Action understanding as inverse planning. Cognition, 113(3), 329–349.
- Berberian, B., Sarrazin, J.-C., Le Blaye, P., & Haggard, P. (2012). Automation technology and sense of control: a window on human agency. PLoS ONE, 7(3), e34075.
- Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.
- Haggard, P., Clark, S., & Kalogeras, J. (2002). Voluntary action and conscious awareness. Nature Neuroscience, 5(4), 382–385.
- Köhler, S., Mecacci, G., & Veluwenkamp, H. (2025). Responsibly engineering control. American Philosophical Quarterly, 62(2), 113–126.
- Himmelreich, J. (2019). Responsibility for killer robots. Ethical Theory and Moral Practice, 22(3), 731–747.
- Levine, S. (2018). Reinforcement learning and control as probabilistic inference: tutorial and review. arXiv:1805.00909.
- Limerick, H., Coyle, D., & Moore, J. W. (2014). The experience of agency in human–computer interactions: a review. Frontiers in Human Neuroscience, 8, 643.
- Matthias, A. (2004). The responsibility gap: ascribing responsibility for the actions of learning automata. Ethics and Information Technology, 6(3), 175–183.
- Mathys, C., Daunizeau, J., Friston, K. J., & Stephan, K. E. (2011). A Bayesian foundation for individual learning under uncertainty. Frontiers in Human Neuroscience, 5, 39.
- Mecacci, G., & Santoni de Sio, F. (2020). Meaningful human control as reason-responsiveness: the case of dual-mode vehicles. Ethics and Information Technology, 22, 103–115.
- Murray-Smith, R., Williamson, J. H., & Stein, S. (2025). Active inference and human–computer interaction. ACM Transactions on Computer-Human Interaction, 32(6), 63.
- Parr, T., Pezzulo, G., & Friston, K. J. (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press.
- Santoni de Sio, F., & Mecacci, G. (2021). Four responsibility gaps with artificial intelligence: why they matter and how to address them. Philosophy & Technology, 34, 1057–1084.
- Santoni de Sio, F., & van den Hoven, J. (2018). Meaningful human control over autonomous systems: a philosophical account. Frontiers in Robotics and AI, 5, 15.
- Sheridan, T. B., & Verplank, W. L. (1978). Human and computer control of undersea teleoperators. MIT Man-Machine Systems Laboratory.
- Tapal, A., Oren, E., Dar, R., & Eitam, B. (2017). The Sense of Agency Scale. Frontiers in Psychology, 8, 1552.
- Wen, W., Yamashita, A., & Asama, H. (2015). The sense of agency during continuous action: performance is more important than action-feedback association. PLoS ONE, 10(4), e0125226.
- Wen, W., Yun, S., Yamashita, A., Northcutt, B. D., & Asama, H. (2021). Deceleration assistance mitigated the trade-off between sense of agency and driving performance. Frontiers in Psychology, 12, 626540.