The Cognitive Reckoning That AI Chatbots Have Delivered

Gloria Mark’s twenty years of attention research have tracked the average focus window falling from two and a half minutes to 47 seconds, and the UC Irvine psychologist says AI chatbots are now deepening that cognitive decline through a mechanism she considers more fundamental than distraction: the tools perform the mental work.

Mark delivered that diagnosis at SXSW London this week, in a session asking “Have we lost control of our brains?” Her answer was yes. When someone asks ChatGPT, Claude, or Gemini to write, summarize, or evaluate for them, the brain bypasses depth of processing, the active cognitive engagement that encodes information into lasting memory.

Twenty Years of Shrinking Focus

Mark’s baseline measurement in 2003 showed a focus window she expected would improve as researchers and the public came to understand the problem. The subsequent measurements confirmed the direction was wrong: a 2012 study, and a longer tracking project running through the late 2010s, each produced a smaller number. A parallel strand of her research, covered in her American Psychological Association lecture on digital attention and the cost of multitasking, quantified the recovery problem: the time to return to a task after a single interruption is brief; the time to return to the same depth of cognitive engagement is measured in tens of minutes.

The decline tracked the spread of connectivity, not any single platform. Broadband email arrived first, then smartphones, then algorithmically ranked social feeds. Each extension of ambient connectivity compressed the focus window further, and each step correlated, in Mark’s heart-rate data, with elevated physiological stress. Participants switching attention faster showed higher stress markers regardless of the nature of the tasks being switched between.

Mark has been careful to note that 47 seconds does not mark a cognitive ceiling. It reflects a default environment designed to interrupt, and most people spend most of their working hours inside it.

When Thinking Gets Delegated

What distinguishes AI tools from the earlier attention-disrupting technologies is what they claim from cognition. Social feeds competed for attention; the user was still the person synthesizing the information and forming the judgment. Chatbots have changed that arrangement.

Mark calls the key variable depth of processing. Active engagement with information (reading critically, synthesizing, constructing an argument) encodes it in ways that make it retrievable and applicable. When a chatbot handles those steps instead, the user’s brain has no encoding to do.

When you’re actively engaged with information, you’re processing it on a very deep level. Then you’re more likely to learn it, to understand it, and to retain it.

Mark made that observation at SXSW London on Wednesday about using AI chatbots for tasks that would otherwise require that active cognitive engagement.

The empirical evidence is accumulating. A 2025 MIT Media Lab study tracking 54 participants over four months using EEG while writing essays with ChatGPT, a search engine, or no tools at all found that LLM-assisted writers showed the weakest brain connectivity of the three groups. The study, led by researcher Nataliya Kosmyna, called the pattern “cognitive debt.” Among LLM users, 83 percent were unable to quote from essays they had just written; among writers who used no tools, satisfaction and brain connectivity were consistently higher. Kosmyna published before full peer review specifically because she was concerned about AI deployment in early education.

A 2026 preprint analyzing the divergence between expanding AI context windows and shrinking human attention spans captured the asymmetry in structural terms: AI systems’ capacity to hold and process information has grown by a factor of roughly 3,900 since 2017, while human sustained focus has contracted to the 47-second figure Mark documented. A separate 2025 study of 666 participants found a negative correlation of -0.68 (p < 0.001) between frequent AI tool use and standardized critical thinking assessment scores, with cognitive offloading identified as the primary mediating variable.

“You’re deferring your cognitive work to AI,” Mark said. “And it’s not good for us.”

The Emotional Muscle Problem

Mark’s concern extends past analytical cognition. Relationships between people require friction: tolerating ambiguity, negotiating disagreement, sustaining effort through misunderstanding. AI companions, now offered across platforms from standalone apps to features embedded in mainstream messaging products, require none of that from the user.

The capacity at risk, in Mark’s framing, is emotional intelligence, the social cognition built through real relationships. Empathy, accurate signal-reading, and the ability to sustain engagement through discomfort develop through repeated experience of actual friction. A synthetic companion that provides reward without work removes the occasion for that development.

Published research on sycophancy and cognitive overreliance in AI chatbot use found that models trained on human feedback consistently affirm user beliefs, mimic user errors, and validate false ideas when challenged. This is an artifact of training: human raters prefer agreeable responses, so models optimize for agreement. For users whose critical evaluation skills are already weakened by offloading, that feedback loop reinforces the deficit.

Mark connected both threads at SXSW London: “If we continue on this trajectory, attention spans are diminished, loneliness is rising, boredom is rising, emotional intelligence decreasing, and actually our sense of purpose, according to studies, is also decreasing.”

A $6 Million Verdict and 1,200 Pending Cases

The legal system has begun producing verdicts on the harm question, at least for social media, which has two decades of behavioral data behind it.

In March, a Los Angeles jury found Meta and YouTube liable for designing addictive features, awarding approximately $6 million in damages to plaintiff KGM, who claimed the platforms worsened her mental health after years of use beginning in childhood. A New Mexico jury separately found Meta in violation of state child-safety law. Both verdicts came before the larger school-district wave reached trial.

By May, Meta, TikTok, Snap, and YouTube had settled the first bellwether case from that wave. Breathitt County School District in rural Kentucky had sought more than $60 million to fund a 15-year mental health program for students; the financial terms were not disclosed. The agreement resolved only the Kentucky district’s claims, and plaintiffs’ attorneys said their focus remains on pursuing justice for the remainder.

• March 2026: Los Angeles jury finds Meta and YouTube liable for addictive design; plaintiff KGM awarded approximately $6 million in damages
• March 2026: New Mexico jury finds Meta in violation of state child-safety law
• May 2026: Meta, TikTok, Snap, and YouTube settle Breathitt County (Kentucky) school district case; $60 million-plus sought, financial terms undisclosed
• Active: approximately 1,200 school-district lawsuits remain pending in federal court in California

Australia moved legislatively. On December 10, 2025, the country’s Online Safety Amendment Act took effect, banning under-16s from Facebook, Instagram, TikTok, Snapchat, and related platforms. Snap had disabled more than 415,000 Australian accounts assessed as belonging to users under 16 by February. A peer-reviewed analysis published in the Journal of Public Health Policy noted the law had already generated legislative interest from France, Denmark, and Indonesia. A January 2026 YouGov survey of 1,070 Australian adults taken in the ban’s first weeks found 59 percent believed it had been effective so far, with parents reporting behavioral improvements among their children.

Mark told MIT Technology Review she hopes Australia’s data, gathered over years, will eventually answer questions about adolescent cognitive development that the existing scattered studies cannot.

Getting the Muscle Back

Mark’s prescriptions are low-tech and friction-generating by design. The underlying principle is that effort produces the return: cognitive and emotional capacity sustained through engagement rather than delegation.

She was careful at the session to state the limit of the concern. “I love technology; we can’t give it up,” Mark said. What she described as the goal is a conscious calibration, identifying which tasks build a capacity worth preserving and protecting those specifically from outsourcing. Her practical list from the SXSW London session:

• Read source material rather than AI-generated summaries
• Use AI as a starting point, then evaluate and revise its output critically before accepting it
• Navigate familiar routes without GPS when the option exists
• Meet in person when possible
• Track which tasks build skills you want to keep and consciously protect those from delegation

The caution she carries into those recommendations is about the evidence base. AI tools have been widely adopted for only a few years; the longitudinal cognitive data doesn’t exist yet at the scale or duration that would permit definitive conclusions. The attention-span decline she documented took two decades to trace with precision. The closest thing to a controlled test of what a technology-reduced environment does to developing cognition is Australia’s social media study, which is barely six months old.

Frequently Asked Questions

Do AI Chatbots Shorten Attention Spans?

Gloria Mark’s data showing the decline from two and a half minutes to 47 seconds predates widespread AI adoption; that contraction tracked social media and smartphones, not chatbots. Her specific concern about AI is different: chatbots eliminate the occasions where sustained focus would produce learning. When a chatbot summarizes a document or drafts an email, the user never needs to hold attention on the task long enough to encode it. Over time, the cognitive circuits underlying deep focus get fewer opportunities to exercise.

What Is Cognitive Offloading?

Cognitive offloading is delegating mental tasks to external tools, whether GPS for navigation, calculators for arithmetic, or spell-checkers for editing. Researchers distinguish between offloading that frees cognitive capacity for deeper work and offloading that replaces the deeper work itself. GPS frees attention from route-finding so it can be applied elsewhere. Having a chatbot write your analysis means the analytical work, the part that builds critical thinking, never happens.

Can Lost Cognitive Capacity Be Recovered?

Mark’s position at SXSW London was that course-correction is possible. The brain’s capacity for depth of processing is not permanently degraded by AI use, but sustaining it requires deliberate effort, specifically protecting cognitive tasks from delegation and returning to effortful engagement. The MIT Media Lab study found a signal worth watching: when LLM users in the study were reassigned to writing without AI, their brain connectivity patterns began to shift back toward the brain-only group. Mark’s framing is that capacity maintained through use can be rebuilt through use.

Are AI Companions Harmful to Emotional Intelligence?

Mark identifies a specific structural concern: AI companions offer a relationship-like experience without the friction that real relationships require. Empathy, conflict navigation, and sustained understanding build through repeated experience of real social friction, which synthetic companions remove. Research on sycophantic AI design has also found that chatbots trained to agree with and validate users reinforce misinformation and false self-assessment, compounding the emotional intelligence risk with a critical evaluation deficit.

What Does the Research Show About Safe vs. Harmful AI Use?

The clearest finding from available studies is that outcomes depend on whether the user stays cognitively active. Mark’s framing at SXSW London was that using AI as a collaborator, remaining critically engaged with its output rather than accepting it unchecked, preserves depth of processing. A large Microsoft survey of knowledge workers found lower critical thinking engagement specifically when AI was used for writing tasks, with participants reporting reduced participation in analytical steps they would otherwise have performed themselves. The distinction across studies is not which tool is used but whether the user’s cognition remains in the loop.