Research programme

The research question

Language and voice models can now pass as human in conversation, and AI systems increasingly act for absent users and operate vehicles and robots near people who never chose to interact with them. The AXI paper argues that many of the resulting failures are not failures of intelligence. They are failures in the layer between capability and people: undisclosed identity, actions taken without current consent, flattery that wins approval while losing accuracy, and machines whose intent bystanders cannot read.

AXI proposes engineering that layer explicitly, as a runtime control plane called the AXI Bridge, governed by one rule: human-grade naturalness, machine-grade honesty. The programme asks whether such a layer can be specified precisely enough to implement and test, and whether its measures track what people actually experience.

Parts of the programme

Scope and relation to existing work

AXI covers three regimes: conversational systems (chat and voice assistants), agentic systems acting under delegated authority, and autonomous systems in shared physical space. It distinguishes four human roles: the principal who delegates, the supervisor who monitors, the occupant who rides along, and the bystander who cannot consent.

The paper positions AXI against design guidance (Microsoft's Guidelines for Human-AI Interaction, Google's People + AI Guidebook), organisational frameworks (NIST AI RMF, ISO/IEC 42001), driving-automation instruments (SAE J3016, UN GTR 26 / R185), the Agent Payments Protocol, Simplex runtime assurance and LLM guardrail libraries. Each of these does things AXI does not, notably certifying driving safety and enforcing content policy. The proposed contribution is to integrate runtime enforcement, a quantified score, non-compensable gates, multi-session trust, delegated mandates and non-consenting humans in one layer (paper §5, Table 9).

AXI is not a certification scheme, does not certify the safety of a vehicle or robot, and adds to independent safety, fairness and regulatory review rather than replacing it.

How AXI relates to other fields

AXI sits where several established fields meet. It borrows from each, and differs in one consistent way: it turns their findings into requirements that are enforced while the system runs and checked with tests (paper §2).

Human–AI interaction (HAI)

Guidelines such as Amershi et al.'s 18 guidelines for human–AI interaction and Google's People + AI Guidebook shape AXI's nine principles. Those guidelines are applied at design review; AXI enforces the same commitments at runtime and measures them.

Human–computer interaction (HCI)

HCI established consistency, feedback and recovery, and it optimizes usability. AXI keeps those principles but optimizesexperience quality: a gated composite of comfort, pacing, trust, repair, continuity and consent.

Human–technology interaction and human-centered AI

Human-centered AI asks that technology serve the people who use it. AXI makes that operational: concrete obligations (consent before action, honesty about identity, repair after errors) with metrics and black-box tests.

The human–AI bridge

The AXI Bridge is the connection layer between AI and people: a runtime control plane between any AI capability and human cognitive bandwidth. It paces output to the person, gates actions on consent, and keeps the system honest about what it is.

Human factors and human–automation interaction

Research on trust in automation (trust calibrated to capability), automation complacency and takeover time underlies AXI's overtrust guard, supervisor attention telemetry and paced escalation for semi-automated vehicles.

Human–robot interaction (HRI)

HRI's distinction between predictable and legible motion becomes AXI's Bystander Legibility measure, which asks whether a robot or vehicle signals its intent to the people around it before it acts.

Conversational AI and affective computing

Humanlike dialogue, natural turn-taking and the uncanny valley motivate the governing rule: be as natural as a person, and as honest as a machine. The Deception Guardrails govern disclosure of AI identity.

AI agents and delegated authority

For agents that act on a person's behalf, AXI specifies signed mandate envelopes, consent records and a rollback protocol, alongside protocols such as MCP, Agent2Agent and the Agent Payments Protocol.

AI safety, ethics and governance

AXI's consent and integrity measures map onto the NIST AI Risk Management Framework, the EU AI Act's transparency duties and ISO/IEC 42001. It adds to independent safety review and is not a certification.

Open problems

1. The naturalness–disclosure trade-off: how to design disclosure so it informs without degrading the interaction.
2. Calibration across cultures and contexts: target bands, turn-taking targets and legibility thresholds across languages, road cultures and homes.
3. Role transitions and multi-agent mandates: people who change roles mid-episode; mandates composed across agents that delegate to agents.
4. Continuity without surveillance: remembering commitments without enlarging the personal-data risk surface.
5. Adversarial robustness and safety cases: gaming of the gates, and verifiable safety arguments for agents acting while principals are absent.

Source: paper §7.3. These are the questions on which outside work would be most useful.

Taking part

• Apply the twelve conformance tests to your own system and share what you find.
• Propose new conformance or guardrail test cases, or collaborate on a validation study: write to srivardhan@sripto.tech.