She spent 9 hours a day generating AI images until reality slipped: Former AI startup executive reveals how AI messed with her sense of normal

In late December 2025, a former AI‑startup executive’s confession of spending nine hours a day generating images with early diffusion models sparked a national conversation about the cognitive side effects of generative visual tools. The story—often framed as “AI psychosis”—reveals a new class of risk that sits at the intersection of human cognition, product design, and regulatory compliance. For technology leaders who rely on image‑generation engines such as Gemini 1.5, Claude 3.5 Sonnet, or GPT‑4o for marketing, design, and content creation, this phenomenon is not merely anecdotal; it demands a strategic response that balances creative freedom with employee well‑being and legal responsibility.

• Risk Identification: Repetitive exposure to imperfect generative images can distort body schema and blur the line between virtual and real imagery, leading to symptoms of AI‑induced body‑image distortion.

• Technology Gap: 2025 models have reduced anatomical errors but still fall short on fidelity metrics (Gemini 1.5 scores < 0.72/1.00 on a “body‑schema fidelity” benchmark).

• Business Impact: High‑intensity creative roles (>8 h/day) are 3.4× more likely to report AI psychosis symptoms, translating into higher absenteeism, lower productivity, and potential legal exposure.

• MHRA guidance (Jan 2025) and a draft DSM‑5‑V addendum (Feb 2025) signal that compliance will soon become mandatory for any product enabling human image generation.

• Companies can differentiate by embedding mental‑health safeguards, tiered access, and cross‑functional teams into their generative AI workflows, creating a new line of “human‑centered” products.

The core of the issue lies in how diffusion models learn to generate human figures. Early 2023‑era models were trained on massive, uncurated internet datasets that included a disproportionate share of hyper‑stylized or anatomically flawed images. Even after fine‑tuning, these models retained subtle biases:

• Over‑smoothing of joint articulation leading to elongated limbs.

• Inconsistent limb ratios when prompted for “slim” figures.

• Recurrent hallucinations of missing or duplicated fingers.

When a user, like Caitlin Ner, spends nine hours daily prompting these models, the brain’s visual cortex is repeatedly exposed to these distorted representations. Neurocognitive research from 2025 indicates that sustained exposure can recalibrate perceptual expectations—a phenomenon analogous to the “mirror neuron” effect but amplified by machine‑generated imagery.

A recent survey of 1,200 creative professionals (WSJ, Dec 23 2025) found:

• Incidence Rate: 12.7% reported symptoms consistent with AI psychosis.

• Workers exceeding eight hours per day on generative tools were 3.4× more likely to report symptoms.

• Estimated average cost of lost productivity and medical intervention is $18,400 per employee annually.

For a mid‑size design firm employing 50 creatives, this translates into an annual hidden cost of roughly $920k—comparable to the budget for a new AI platform. Ignoring the risk could also expose firms to civil liability if employees develop diagnosable conditions linked to workplace practices.

The UK Medicines and Healthcare products Regulatory Agency (MHRA) issued a guidance note in January 2025 recommending “mandatory safety disclosures for any product that allows users to generate human imagery.” Similar notices are pending from the EU’s European Medicines Agency (EMA) and the U.S. Food & Drug Administration (FDA). Key compliance points include:

• Explicit disclosure of potential cognitive side effects.

• Implementation of “break reminders” after a set number of continuous prompts.

• Optional “reality check” filters that flag images with extreme body ratios or unrealistic features.

Non‑compliance could result in fines up to £2 million for UK firms and market withdrawal for EU/US products. Companies already integrating Gemini 1.5’s Reality Check mode and Claude 3.5 Sonnet’s Real‑World Filter are ahead of the curve, but adoption rates remain low—only 18% of surveyed firms have enabled these safeguards.

While AI psychosis poses a clear risk, it also opens avenues for differentiation:

• The 2025 AI Ethics Summit charter recommends embedding mental‑health metrics into product roadmaps. Firms can lead by offering tools that monitor cognitive load and provide de‑briefing prompts after intensive sessions.

• By segmenting users into “high‑risk” (designers, illustrators) and “low‑risk” (hobbyists), companies can deploy stricter safeguards for the former while preserving creative freedom for the latter. This approach is already adopted by a few boutique studios that report a 27% reduction in reported symptoms.

• As the American Psychiatric Association drafts DSM‑5‑V addenda, insurers are poised to cover AI‑related mental health claims. Early adopters of safety protocols can negotiate lower premiums and attract clients seeking compliant solutions.

Below is a step‑by‑step framework that aligns technical deployment with regulatory compliance and employee well‑being:

• Audit Existing Workflows: Map out all touchpoints where employees interact with generative image tools. Identify peak usage periods and potential exposure thresholds.

• Select Model & Safeguards: Opt for the latest Gemini 1.5 or Claude 3.5 Sonnet releases that include Reality Check or Real‑World Filter options. Verify that your deployment environment supports real‑time flagging.

• Embed Break Reminders: Integrate a system that prompts users to take a 5–10 minute break after every 90 minutes of continuous image generation. Leverage existing productivity suites (e.g., Microsoft Teams or Slack bots) for reminders.

• Develop Self‑Assessment Checklists: Create short, daily questionnaires that prompt employees to rate visual fatigue and body‑image distortion symptoms. Store responses in a secure analytics dashboard.

• Train Cross‑Functional Teams: Assemble a squad comprising ML engineers, UX researchers, psychiatrists, and HR specialists to oversee tool governance. This team should review model outputs quarterly for bias and fidelity drift.

• Document Compliance: Maintain logs of safety disclosures, user consent forms, and break reminder metrics. These records will be essential if regulatory audits arise.

• Iterate Based on Feedback: Use the analytics dashboard to identify patterns (e.g., certain prompts that consistently trigger high distortion scores) and adjust model fine‑tuning or prompt guidelines accordingly.

A pilot study conducted by a leading ad agency in 2025 demonstrated that implementing the above framework reduced reported AI psychosis symptoms by 68% over six months. The cost savings were quantified as follows:

• $42,000 per employee annually.

• $35,000 per employee (average replacement cost).

• Estimated potential fine avoidance of £1.2 million across the firm.

Total annual benefit: approximately $77k per employee. When scaled to a workforce of 200 creatives, the organization stands to save roughly $15.4M annually—well above the upfront investment in tool upgrades ($350k) and training ($120k).

By 2027, we anticipate a shift from reactive safeguards to predictive analytics:

• Models that estimate risk scores based on prompt complexity and user history.

• Interfaces that adjust image fidelity or suggest alternate prompts when distortion likelihood exceeds a threshold.

• Adoption of the “Body‑Schema Fidelity” score as part of product certification, similar to how GPU benchmarks are currently used.

Companies that invest early in these capabilities will not only mitigate risk but also position themselves as leaders in ethical AI design—a growing demand among investors and consumers alike.

• Conduct a company‑wide audit of generative image usage hours and identify high‑risk roles.

• Deploy the latest Gemini 1.5 or Claude 3.5 Sonnet releases with built‑in safety modes.

• Automate reminders after 90 minutes of continuous use.

• Form a team that includes mental‑health professionals to oversee tool design and usage policies.

• Create dashboards for cognitive load, distortion scores, and compliance logs.

• Proactively submit safety disclosures and seek early feedback on upcoming guidelines.

• Offer training modules that explain the risks of AI psychosis and how to recognize symptoms.

• Negotiate coverage for AI‑related mental health claims as a value‑add for clients.

By treating generative visual tools not just as creative accelerators but also as complex systems with human‑centered impacts, enterprises can safeguard their workforce, comply with emerging regulations, and unlock new market opportunities in 2025 and beyond.