Artificial intelligence arrived in our public sphere with many lofty promises. AI would automate the boring stuff, free humans to do the meaningful work, and maybe even help us make fairer, more rational decisions. What it actually delivered was more complex. AI systems now filter what people see, shape what they believe, and increasingly decide who gets hired, who gets a loan, and who is flagged as a risk. It stands out that both the positive and negative effects are social problems, not technical ones. Yet we fail to admit that once software begins to steer everyday life at scale, it ceases to be just a tool and becomes an institution.
The core misconception is the belief that more AI will solve problems like polarization, misinformation, bias, or job loss. When things go wrong, tech companies want to add more data and layers. However, these problems stem from incentives, power dynamics, and a lack of transparency. Approaching social fallout as mere optimization ensures the next systems will inherit the same problems, only faster and on a greater scale.
Focusing on a few fundamental questions, we can already see that technology isn’t the issue. After all, the answer to: “Who sets the objectives?”, “Who controls the data?” and “Who suffers the impact of errors?” are never AI. Yet, clarifying these points reveals that social outcomes are decided well before technology is built.
Bias In, Bias Out – Now With An AI Confidence Score
Silicon Valley promises that AI is the antidote to human prejudice. With enough data, machines crunch the numbers, and subjectivity will give way to clean, impartial decisions. Unfortunately, models trained on historical data faithfully learn past discrimination and then reproduce it in a polished, automated form. Systems used in hiring, lending, or policing have repeatedly shown to treat certain groups more harshly because they were trained on biased records. The cynic might say: “Bias has not disappeared. We simply added an API.”
Despite layering more technology, underlying structural issues persist. Addressing fairness requires more than technical fixes. We must confront deeper social problems rather than hope that automation alone will solve the issue them.
Bias Scores When Detecting AI
Let’s look at education, especially at the trend of using AI to detect students cheating. Today, you can find a number of AI detection tools that give you the probability of a student using AI. The tools can even go so far as to mark the sentences likely to be AI-generated and, funnily enough, create a cheating note for the instructor to pass on. Yet, none of the data takes into account the student’s background. Their “moderate confidence in AI generation” doesn’t account for the fact that foreign students are more likely to use rigid grammar structures.
Yet by outsourcing detection and even accusations to AI, educational organizations and society can avoid confronting the question of education’s value. Questions like “Is our Teaching Relevant?” or “Is everyone represented in the debate?” remain unasked and unanswered. AI can make up confidence intervals and correlate styles. Yet, it cannot replace public debate, legal rights, and institutional accountability.
Social Fabric Is Not A Data Problem
Yet, it is convenient to blame social problems on AI. Image generation tools and AI social media recommendation engines did not cause our fractured society. AI only made it cheaper and faster to produce convincing fakes, and harder to ignore. Likewise, automation did not start the conversation about the future of work. It condensed decades of labor disruption into a few product cycles. What changed is not human nature but the speed, scale, and opacity with which our existing tensions are amplified.
Trying to fix these dynamics with more AI misses the point. Our social trust depends on shared norms, stable institutions, and mechanisms to challenge power. Those emerge from law, culture, and politics, not algorithms. AI tools can flag or filter, but they can’t rebuild trust or navigate value trade-offs that require human accountability and decision-making.
A Seat At The Table
Google’s diverse Nazis still illustrate what happens when corporate politics and technology miss the mark. As a reminder, when Google tried to ensure its image generation tool would show a diverse society, it falsified historical images to create a very diverse 3rd Reich army. Ultimately, they had to admit they tried to falsify history to appease today’s politics.
Yet, at no point did their attempts challenge the underlying problem, that skin color in today’s world still determines access. Access that includes the agency to decide what should be acceptable in technology. Yet, they tried to reframe the political debate as a technical issue. Thus, it became a problem they could solve. Unfortunately, not only did their AI get things wrong, but the system even tried to redefine what right and wrong look like. In the end, we all got a front-row seat to what happens when social problems turn into bug numbers in an issue tracker.
Choosing Governance Over AI Autopilots
The uncomfortable truth is that AI is excellent at making existing power structures more efficient. It routes attention toward what keeps users engaged, optimizes logistics in ways that often sideline workers, and extracts patterns from personal data that are profitable precisely because they are intimate. Left on autopilot, these systems will not steer toward fairness or inclusion. They will steer us towards whatever metric their owners reward. Expecting those same systems to self-correct the social imbalances they exploit is like expecting a high-frequency trading algorithm to design a more equitable economy.
This does not mean rejecting AI outright. It means refusing the comforting myth that every social externality can be “handled in the stack.” The real leverage lies outside the model. We need regulations that set boundaries, procurement policies that demand transparency, and education that equips people to question automated decisions. We need institutions willing to say that some functions, including sentencing, asylum decisions, and existential medical triage, are simply too bound up with human dignity to be delegated to statistical guesses, no matter how accurate they appear in testing.
AI will continue to shape society, but it cannot absolve society of the responsibility to shape AI. The systems built today will encode answers to questions about whose time matters, whose privacy counts, and whose errors are tolerable. Those answers cannot be left to an optimization routine. They belong in parliaments, boardrooms, union halls, and living rooms. The temptation to let another model clean up the mess is strong, especially when quarterly results are on the line. Resisting that temptation is the first real test of digital maturity.
Leave a Reply