Manifesto for real AI and algorithmic transparency and openness

Transparency is regularly cited as a core principle of ethical, responsible and trustworthy AI. 

However, rhetoric and reality are often poles apart, with transparency approached in a partial, piecemeal and reactive manner. 

Here's why real AI and algorithmic transparency and openness is needed, and what it should look like.

For better or worse, AI, algorithms and automation are here to stay

• Fundamental to many people’s daily lives, shaping their experiences, beliefs, and behaviours

• Increasingly central to the daily workings of government, companies, media, and other organisations

• Reflect the people who design, build, and manage them

The risks of AI, algorithmic and automation are increasing

• Systems are becoming larger and more complex

• Fragmented standards, little legislation

• Understanding, confidence and trust in AI, algorithms and automation are low

• Expectations of fair, transparent, and accountable systems are rising

• Scrutiny by legislators, regulators, NGOs, and the media is increasing

• Societal, environmental and reputational risks are escalating

Many AI, algorithmic and automation systems are opaque and unaccountable

• Typically, users have little or no idea how the systems they are using or are being assessed, nudged, coerced or manipulated by work 

• Equally, they may not know they are interacting with AI, algorithmic and automation systems at all

• The actual or potential workings, limitations, risks and impacts of AI, algorithmic and automation systems are often concealed, disguised or deflected

• Many AI and algorithmic systems are unavailable for meaningful external analysis or investigation

AI, algorithmic and automation transparency efforts lack substance

• User information tends to be narrow, skin-deep and piecemeal

• Feedback, complaint and appeal systems are slow and partial

• Remain largely the preserve of technology experts

Transparency and openness are two-edged swords

• Opportunity to build confidence, esteem, and trust

• Increases visibility and expectations whilst exposing AI systems to manipulation, fraud, legal liability, and loss of IP

• Facilitates and shapes AI and algorithmic governance and ethical decision-making - for better or worse

• Requires compromise/trade-offs.

Meaningful

• Provide clear, concise, accurate, timely, and usable information 

• Consistent across all channels and to all audiences 

• Visible at the first point of contact with the system, and across all relevant touchpoints

• Responsive to feedback, complaints and appeals

• Honest, including during incidents, crises and controversies

Verifiable 

• Provide reasonable third-party expert access to data, code and model 

• Make third-party audit and investigatory reports public

Inclusive

• Involve users and other relevant stakeholders in system design, development and deployment

• Provide formal feedback loops to relevant audiences, and share feedback as widely as possible

Accessible

• To all relevant users, including the disabled, the elderly, non-internet users, and other disadvantaged people.

Updated: October 9, 2024