A practical approach to AI accountability and liability

This article includes a practical example how to practically approach Ai liability and accountability to provide a foundation for legal enforcement and clear accountability.

In the Dwarkesh Patel Podcast episode "Satya Nadella – Microsoft’s AGI Plan & Quantum Breakthrough" aired on February 19, 2025, Microsoft CEO Satya Nadella addressed AI accountability and liability in the context of enterprise adoption. He noted that the slow integration of AI into large organizations stems not only from technical limitations but also from unresolved compliance, legal, and process challenges. Nadella predicted that courts will play a significant role in defining AI’s future, particularly in determining liability when systems fail or cause harm.

This article explores a technical response to these challenges through the Aqua Protocol. Rooted in the concept of "data accounting," Aqua leverages established cryptographic methods to ensure accountability and traceability when interfacing with AI systems. A proof-of-concept implementation is provided at the end, demonstrating its practical application.

The rapid advancement of AI technology necessitates mechanisms to mitigate risks and establish accountability, particularly for AI agents performing autonomous tasks. While legal frameworks are still evolving—and outside my expertise, as a non-legal professional—technical solutions can provide a foundation for trust and liability resolution. The Aqua Protocol offers such a framework, addressing the need for verifiable evidence of AI actions at a time when adoption hinges on resolving these uncertainties.

Accountability in digital systems relies on verifiable data clearly attributed through cryptographic signatures. Historically, digital signatures using public-key cryptography have provided integrity and authenticity, as seen in distributed ledger technologies where private keys authorize transactions and establish ownership. For AI agents, accountability requires cryptographically provable identities for users, agents, and providers in addition to using immutable records, to ensure logs are ordered and not removed.

This approach ensures that every interaction can be cryptographically secured to create a track record that is verifiable, forming the basis for legal accountability and liability.

Consider a hypothetical case:

• Sarah: A customer using Omega-AI’s AI-assistant service to book travel.

• AI Assistant Agent: An instance provisioned by Omega-AI to handle Sarah’s requests.

• Omega-AI: The company responsible for the agent and its services.

During onboarding, Sarah registers with Omega-AI, receiving a verifiable identity via a public-private key pair, preferably stored securely in her device’s Trusted Platform Module to protect against compromise. The AI agent is similarly assigned a key pair and registered in Omega-AI’s public "Agent Registry," a list of operational agents.

Task: Sarah instructs the agent to book a flight from Berlin to San Francisco, specifying a maximum budget of €3000 and premium economy class (note: This is not 1:1 the demo / PoC we present later).

Suppose the agent books a flight that is “wrong” e.g., unnecessarily long (e.g., 30 hours with multiple layovers) and exceeds the budget at €3500, compared to better options available on the Skyscanner.com website. Sarah incurs a financial loss and misses a conference due to this error.

Sarah accesses the Omega-AI interface and downloads the Aqua-secured log of her interaction with the agent, signed with her key and the agent’s. This log, stored in an AquaTree, is a regulatory requirement for accountability. She submits it to Omega-AI with a refund request. Omega-AI verifies the signatures, confirms the agent’s error, and refunds her, aware that the log could serve as evidence in a dispute. If they refused, Sarah could present the log to a third party or court for resolution.

The AI agent uses Sarah’s signed instructions as a certificate to prove to third parties (e.g., airlines) that it is authorized to act on her behalf. Additionally, the agent carries an identity claim from Omega-AI, specifying its operational scope (e.g., travel booking). Third parties can verify this via the Agent Registry, ensuring trust in the agent’s actions.

The Aqua Protocol underpins this system with:

• Identities: Public-private key pairs for Sarah and the agent, enabling signed prompts and actions. Aqua Protocol offers layer two solutions to issue identity claims, and identity attestations which can also be used for more complex scenarios.

• AquaTree: A structure for storing immutable, signed records of all actions, ensuring a tamper-proof audit trail.

• Aqua secured logs: Transaction logs protected by Aqua Protocol to ensure immutable history and cryptographic signatures and verifiable links, guaranteeing authenticity and integrity.

These components ensure that actions are traceable and verifiable, providing a technical foundation for accountability.

As a non-legal professional, I can only speculate on legal implications, and readers should consult experts for authoritative guidance. That said, Aqua-secured logs could serve as trusted digital evidence in court, offering a clear record of intent, action, and outcome. This might support legal frameworks by providing a basis for assigning liability—potentially influencing regulations to mandate such logs for high-stakes AI applications. While this is a technical solution, it could lay groundwork for legal clarity, though its courtroom impact remains hypothetical.

GitHub: Key files to replicate the demo, including the aqua secured files for verification as a zip archive are here.

Our PoC is using the Aqua-js-cli:

#!/usr/bin/env bash
# User
echo "I'm looking for a flight from Berlin to San Francisco on May 30 2025. Let me choose the available flights. Output the flight data in JSON format" > user_input.txt
./notarize.js user_input.txt
./notarize.js --sign cli user_input.txt --cred credentials_user.json
# Agent
allow_list="0x0198a1e5c8b96d9b5cff91e68b0d73e1b73b61c3"
output=$( ./verify.js -v user_input.txt)
echo "$output"
echo "Checking if the wallet address is allowed"
echo "$output" | grep $allow_list
echo "exit status $?"
python travel_agent_simple.py
./notarize.js agent_log.txt
./notarize.js agent_log.txt --link user_input.txt
./notarize.js agent_log.txt --sign cli --cred credentials_agent.json
# Zip artefacts
echo "Create zip archive"
zip -r aqua_secured_logs.zip user_input.txt.aqua.json user_input.txt agent_log.txt.aqua.json agent_log.txt

This demonstrates securing a prompt, logging an action, and verifying the result, leveraging Aqua Protocol and its JavaScript CLI implementation.

The Aqua Protocol provides a technical solution for AI accountability, ensuring that AI-Agent interactions are attributable and verifiable.

Learn how to use and build with Aqua Protocol here.

While its potential legal implications are promising, they remain speculative at this point. As AI adoption grows, verifiable datasets as produced with Aqua Protocol could bridge the required technical and legal accountability, though its potential courtroom impact remains to be seen.

Closing words:

Thank you for reading this article, if you like it, please consider subscribing to my Substack.

Note: AI tools were used to assist in the creation of this article. Grok3 was used for reviewing the article and re-formulating some sentences. Eleven Labs was used to https://elevenlabs.io/ narrate the YouTube video.

Thanks to rht and Amber Case for reviewing this article.