For consent to be ethically meaningful, it must be informed (people understand what they are agreeing to), freely given (not coerced or buried in terms and conditions), specific (not vague blanket permissions), and revocable (individuals can change their mind and have their data removed). 

Consent in AI is multidimensional, encompassing: 

• Data collection consent. Agreement to the harvesting of personal data (text, images, audio, biometrics) for the training or operation of AI models.

• Inference consent. Agreement to having AI systems draw inferences about an individual from their data (health, behaviour, emotions, identity).

• Output consent. Agreement to one's voice, likeness, or creative style being replicated or synthesised by generative AI.

• Automated decision-making consent. Awareness that one's data is being processed by automated systems making consequential decisions (credit, employment, healthcare, policing).

• Ongoing and withdrawable consent. The ability not merely to consent once but to withdraw consent and have that withdrawal effectively enforced downstream.

Why it matters

Consent is foundational to human autonomy and dignity. Without it, individuals lose meaningful control over their own identities, bodies, and creative outputs. 

At a societal level, widespread non-consensual data use erodes trust in institutions and technology, tilts power heavily toward large corporations and governments, and can entrench systems of exploitation - particularly of those least able to advocate for themselves. 

Consent also underpins the legitimacy of AI-driven decisions: systems built on data taken without consent lack a sound ethical or legal foundation, regardless of how technically sophisticated they are. 

Impacts & consequences

When consent norms break down, the consequences can be wide-ranging:

• Exploitation of creators and workers. Artists, voice actors, writers, and photographers find their work or likeness reproduced and monetised without compensation or credit.

• Social manipulation. Automated systems can use non-consensual data to micro-target individuals with misinformation or predatory advertising. 

• Surveillance and chilling effects. People modify their behaviour (how they speak, where they go, what they write) when they realise they are being captured by systems they never agreed to.

• Discriminatory outcomes. Data taken without consent is often unrepresentative or contextually misused, feeding biased AI outputs that harm marginalised communities.

• Loss of bodily and identity sovereignty. Non-consensual facial recognition, biometric capture, and synthetic voice/image generation reduce individuals to data points they have no power to control.

• Legal liability and regulatory action. Organisations face fines, lawsuits, and reputational damage when consent violations are exposed.

Causes & contributing factors

Common sources of consent failures in AI include:

• Scale and opacity of data scraping. Automated pipelines harvest vast datasets from the public internet, treating "publicly accessible" as synonymous with "consented to."

• Dark patterns in terms of service. Consent is buried in impenetrable legal text, defaulted to opt-in, or bundled with unrelated permissions.

• Absence of universal legal frameworks. Consent rules vary wildly across jurisdictions, creating regulatory arbitrage opportunities.

• Commercial pressure. The competitive need for large, diverse training datasets incentivises cutting consent corners.

• Technological complexity. Individuals often cannot understand what they are consenting to, making meaningful consent difficult even when it is nominally sought.

• Retroactive use. Data collected for one purpose (e.g. a translation dataset) is later repurposed for AI training without fresh consent being sought.