Five ways to ethically build and use AI

Executive overview

As AI scales rapidly, we risk embedding human biases into systems that affect millions of lives. The challenge is ensuring humans remain at the center of AI development and deployment—not just in code, but in governance, safety, user experience, and organizational culture.

The solution: five concrete practices that ground ethical AI in human values of fairness, equality, honesty, and responsibility.

Human-centered AI requires intentional design across five domains: inclusive governance, user safety as culture, skeptical testing, near-term security focus, and honest language about what AI actually is.

Give everyone a say

• Invite commercial and non-commercial voices to strategize together; commercial actors scale R&D, non-commercial actors spot blind spots and refocus on human impact.
• Racial and gender diversity in AI teams is non-negotiable—homogeneous teams encode their own values into systems that affect everyone.
• The demographic makeup of AI builders determines whether AI amplifies humanity's best or worst attributes.
• Programs like AI for All deliberately recruit underserved communities into AI careers, reshaping whose voices shape the technology.
• User feedback matters: as you use AI today, the models learn and improve to better represent you.

Prioritize user safety above all else

• Ask constantly: Who are you actually serving—the end user or the business paying for the system?
• Avoid models funded by advertisers; when the buyer isn't the user, the system serves interests that conflict with user welfare.
• Embed safety into organizational culture, not as a separate checklist; every team must own outcomes, not just a designated safety squad.
• Implement non-bypassable guardrails: Pi won't respond to prompt hacking and maintains firm boundaries while staying respectful.
• Scrutinize what customer data you share with AI systems and be transparent about when and how you use AI.

Question AI's omniscience

• Large language models display slices of intelligence but often fail at tasks humans find trivial—they hallucinate, misunderstand context, and inherit training data bias.
• Prompt hacking and false citations (like fabricated court cases) show AI confidently invents information when uncertain.
• Facial recognition algorithms arrest innocent people because bias in training data produces biased predictions, yet police departments often have no accuracy metrics.
• Managers and teachers must test systems on themselves before deploying them on others; lived experience reveals limitations that abstract metrics hide.
• Productivity monitoring tools encourage gaming the system (moving your mouse to inflate scores) rather than improving actual work.

Address near-term threats over existential ones

• Misinformation at scale can destabilize elections; cyber attacks are 10–100x more likely to exploit AI than disinformation is.
• Generative AI lowers the barrier to entry for both defenders and attackers—we're in an asymmetric period where bad actors have the advantage.
• Red teaming (inviting external hackers to break systems) works: 2,200 participants found vulnerabilities in major LLMs at DEFCON 2023.
• Practical security solutions exist now; they require collaboration between technologists, policymakers, and domain experts—not endless debate about robot apocalypse scenarios.

Reframe how we talk about AI

• Anthropomorphizing AI ("the algorithm replaced teachers") erases human responsibility and absolves developers of accountability.
• When we speak as though AI makes autonomous decisions, we create a self-fulfilling prophecy of human irrelevance and disempower users.
• Developers must clearly remind users that AI is a tool, not an agent with its own intentions—especially as models become more fluid.
• Language shapes imagination; if we linguistically hand agency to machines, we mentally construct a world where they've replaced us.