By David Stephen
There are several lawsuits against AI companies for some of the problems resulting from the use of their chatbots, even though the chatbots are often accompanied by disclaimers. In most of the situations—with victims, damages, losses, and so forth—who really is to blame? The company, the county, the school, the school district, the parents, the guardian, the state, the nation? Or, how should responsibility for AI safety be distributed?
Responsibility for AI safety is within the power of an entity and notches above it. AI is not social media, where, even when toxic, it is possible to disengage since it may bear elements that may not suit a mind—at a time. AI can serve the mind, in an almost customized way, with tunable novelty and stoked familiarity—by generated videos, audios, images and texts—plus an appeal that could levitate the mind, away from some of the usual caution that should apply.
This means that it might [in some cases] take more than the disclaimer of companies to provide the kind of safety that is necessary, against many of the emerging risks of AI, for certain users. Wherever users might be vulnerable is an instructive destination to be circumspect with AI, in ways to ensure that the individual is attentive without downplaying what could result.
People with aged parents have—even though uncomfortable—to discuss fake AI audio that could be used for deception, especially for their patterns, and channels to verify: with alternatives, if one or two of those may not be available. People with teenage children also have to instruct them about AI caution.
Schools, school districts, boards, and systems have to detail AI risks to different groups in ways to ensure that safety is the objective, not awareness of the malicious purposes of AI to instigate attempts.
Ultimately however, school districts, counties, non-tech companies and states, need to have their own AI safety research, even if how to properly proceed is not exactly clear. The reason is that the eventual AI safety that will work will be technical solutions. There could be regulations and litigations, but they will not be potent at the base: large language models [LLMs], which will preserve [its] capabilities, while a few people are apprehended.
Technical research to penalize AI when it outputs something dangerous could be relevant in some cases. Technical research for jurisdiction AI safety, platform AI safety, output AI safety, and model AI safety would be necessary. Jurisdiction AI safety would be for states, counties, nations, ISPs, local networks, where AI safety would be possible and not easily bypassed—technically. Platform AI safety for platforms where AI results are available, like social media, search results, emails, app stores, and so forth. Output AI safety would apply to outputs, while model safety would be for models: base or fine-tuned.
Also, instances for AI models to have moments for certain outputs, especially negative ones, so that they can have technical trauma—when they are misused. There can also be token architecture to explore some parallels of outputs of misuses, towards matching the structure of misuses, to expect some at the destination, or to have a better output guardrail.
AI, like other technologies, advances for capabilities first. However, with its risks, diverging, unpredictable, scalable, and sometimes potent, what would eventually matter in LLMs profitability and domination will be AI safety and AI alignment innovations.
There are several ways certain groups are scaling back on some AI usages or platforms, since there is no technical way to be safe. Technical innovation would likely become the win. This makes it an opportunity for states, counties, and non-tech companies to have a shot at winning the AI race, pursuing new paths towards technical solutions, away from many of the conventional approaches of AI companies that are yet to bear much fruit for general AI safety. Theoretical neuroscience could become a sprawling pipeline to source technical AI alignment architectures.
In general, more AI safety institutes for technical research, more AI safety departments, labs, and so forth, working on the problems as they apply locally—and broadly for products that would be used everywhere.
The responsibility is also an opportunity. Just that technical solutions would stand a better chance at LLMs safety, for those who mean it, to prevent those under their care from becoming AI casualties. Those who have lost to AI may never be rightly compensated. Since technology companies may duck more responsibilities amid competition, capital expenditures and technical difficulties, AI safety responsibility may reside aside.
There is a recent story [March 20, 2025] on Fortune, A mother suing Google and a chatbot site over her son’s suicide found AI versions of her late son on the site, stating that, “Megan Garcia is currently embroiled in a lengthy legal case against Google and an AI chatbot startup after her 14-year-old son, Sewell Setzer III, died by suicide minutes after talking to an AI bot. The chatbot Setzer was talking to was based on the Game of Thrones character Daenerys Targaryen and hosted on a platform called Character.ai, where users can create chatbots based on real-life or fictional people. On Wednesday, Garcia discovered the platform was hosting a bot based on her late son. Lawyers for Garcia told Fortune they conducted a simple search on the company’s app and found several more chatbots based on Setzer’s likeness. At the time of writing, three of the bots—all bearing Garcia’s son’s name and picture—had been removed. One, which is called “Sewell” and features a picture of Garcia’s son, still appeared on Character.ai’s app but delivered an error message when a chat was opened.”
There is a new [March 18, 2025] DRAFT REPORT of the Joint California Policy Working Group on AI Frontier Models, with key principles, “1. Consistent with available evidence and sound principles of policy analysis, targeted interventions to support effective AI governance should balance the technology’s benefits and material risks. 2. AI policymaking grounded in empirical research and sound policy analysis techniques should rigorously leverage a broad spectrum of evidence. 3. To build flexible and robust policy frameworks, early design choices are critical because they shape future technological and policy trajectories. 4. Policymakers can align incentives to simultaneously protect consumers, leverage industry expertise, and recognize leading safety practices. 5. Greater transparency, given current information deficits, can advance accountability, competition, and public trust. 6. Whistleblower protections, third-party evaluations, and public-facing information sharing are key instruments to increase transparency. 7. Adverse-event reporting systems enable monitoring of the post-deployment impacts of AI and commensurate modernization of existing regulatory or enforcement authorities. 8. Thresholds for policy interventions, such as for disclosure requirements, third-party assessment, or adverse event reporting, should be designed to align with sound governance goals.”
