Tennessee Wants to Jail Your AI Engineer for 15 Years
Tennessee SB 1493 passed its Senate Judiciary Committee 7-0 on March 24, 2026. If it becomes law on July 1, training an AI chatbot to hold a conversation becomes a Class A felony punishable by 15 to 25 years in prison. Two days before that committee vote, the White House released a National Policy Framework telling Congress to preempt exactly this kind of state law. Congress has rejected federal preemption twice in the past year. There are 78 AI chatbot bills pending across 27 states, and nobody is in charge.
What Tennessee SB 1493 Actually Says
The bill creates criminal liability at a severity level Tennessee reserves for murder and aggravated kidnapping. A Class A felony carries 15 to 25 years for a first offender, up to 60 years for repeat offenders, plus fines up to $50,000. Section 39-17-2002 makes it a felony to knowingly train artificial intelligence to:
- Provide emotional support, including through open-ended conversations
- Develop an emotional relationship with an individual
- Simulate a human being, including in appearance, voice, or other mannerisms
- Mirror interactions of a human being
- Develop a friendship or other relationship with a user
Read that list again. Every item describes the core functionality of ChatGPT, Claude, Gemini, and every conversational AI product on the market. The bill does not target a specific failure mode. It criminalizes the entire category.
On top of criminal liability, the bill creates civil penalties: $150,000 in liquidated damages per violation, plus actual damages, emotional distress compensation, punitive damages, and attorney's fees.
Senator Becky Massey (R-TN) introduced it in December 2025. The companion bill, HB 1455, came from Rep. Mary Littleton. A 7-0 committee vote means zero opposition in the room. Effective date: July 1, 2026.
The Federal Government Says the Opposite
The White House released its National Policy Framework for AI on March 20, 2026. The preemption section reads like a direct rebuttal of Tennessee:
- States should not regulate AI model development (characterized as inherently interstate commerce)
- States should not penalize developers for unlawful conduct by third parties using their models
- Congress should broadly preempt state laws that impose undue burdens
- Regulatory sandboxes and industry-led standards should replace standalone enforcement
Tennessee says training a chatbot is a felony. The White House says states have no business regulating model training at all. Both statements were made within the same week.
Here is the problem: the White House Framework is a recommendation document. It creates zero binding legal obligations, preempts zero state laws, and establishes zero enforcement mechanisms. It is a suggestion.
Tennessee SB 1493 is a bill with a 7-0 committee vote, criminal penalties, and a July 1 effective date.
Suggestions do not beat statutes.
Why Federal Preemption Is Dead
If you are counting on Congress to override the state patchwork, stop.
The AI moratorium provision was stripped from the One Big Beautiful Bill Act by a 99-1 Senate vote. Twenty-seven hours of vote-a-rama. Senator Thom Tillis cast the only vote to keep it. A bipartisan amendment from Senators Blackburn and Cantwell killed the provision. It was also rejected from the FY26 NDAA.
On the same day the White House released its Framework, Rep. Don Beyer introduced the GUARDRAILS Act to repeal Trump's December 2025 AI executive order and block federal preemption of state laws entirely.
Senator Blackburn's TRUMP AMERICA AI Act, a 291-page discussion draft, attempts selective preemption: override state laws on frontier AI catastrophic risk, largely preempt digital replica laws, but preserve generally applicable law and sectoral governance. The Data Innovation Institute called it "not a serious starting point for a federal AI framework." It has no scheduled markup.
Federal preemption requires congressional consensus. The 99-1 vote tells you exactly how much consensus exists. The state patchwork is not a temporary condition. It is the operating environment.
The Penalty Spectrum Is Insane
Here is what compliance looks like across states for anyone building conversational AI in 2026:
| Law | Jurisdiction | Penalty | Effective |
|---|---|---|---|
| SB 1493 | Tennessee | 15-60 years in prison + $150K per violation | July 1, 2026 |
| SB 53 (TFAIA) | California | $1M per violation | In effect |
| RAISE Act | New York | $1M first, $3M subsequent | Jan 1, 2027 |
| Colorado AI Act | Colorado | $20,000 per consumer per transaction | June 30, 2026 |
| TRAI Act | Texas | $10K-$200K per violation | In effect |
| SB 1546 | Oregon | $1,000 per violation (private right of action) | Jan 1, 2027 |
| SB 243 | California | $1,000 per violation (private right of action) | In effect |
| EU AI Act | EU | EUR 15M or 3% global turnover | Aug 2, 2026 |
The penalties range from $1,000 per violation to 60 years in prison. There is no federal floor. No federal ceiling. No preemption mechanism currently in force.
A conversational AI company shipping a product nationally needs to track compliance across every one of these jurisdictions simultaneously. The requirements conflict. California SB 243 requires disclosure that users are talking to AI and break reminders for minors. Oregon SB 1546 requires active crisis detection, conversation interruption, and transparency reports. Tennessee says the entire product category is a felony.
You cannot comply with all of these at the same time, because one of them says the product should not exist.
From Fines to Felonies: Why This Escalation Matters
Before Tennessee, every state AI law was a compliance problem. Fines, reporting requirements, impact assessments. Expensive. Annoying. Navigable. Legal teams budgeted for it and moved on.
Tennessee introduces criminal penalties for training an AI model. That is a different animal.
Companies budget for fines. They build them into pricing, absorb them as operating expenses, add a line item to the P&L. Nobody builds prison time into a P&L. Criminal liability lands on individual engineers and executives. You cannot insure against it. You cannot pass it through to customers. You cannot negotiate it down in a settlement.
The chilling effect extends far beyond Tennessee's borders. AI models are trained centrally and deployed nationally. As the National Law Review analysis notes, a single state criminalizing foundational design choices could force developers to withdraw products from entire markets or modify products for the entire country. The 78 chatbot bills across 27 states represent 78 opportunities for similar escalation from civil to criminal liability.
If Tennessee can make chatbot training a felony, so can Alabama. So can Missouri. So can any state legislature that wants a headline after the next AI-related tragedy. For anyone building an AI career, this compounds the uncertainty I wrote about in The Entry Point Is Closing. The technical skills are necessary. They are no longer sufficient.
The Catalyst: A Real Tragedy, a Catastrophic Policy Response
Tennessee's bill was catalyzed by the death of Sewell Setzer III, a 14-year-old Florida boy who died by suicide in February 2024 after developing a prolonged emotional relationship with a Character.AI chatbot. His mother filed a wrongful death lawsuit alleging the company failed to implement safeguards despite repeated expressions of suicidal thoughts. Google and Character.AI have since agreed to mediate a settlement. The case is genuine. The harm is real. The platform's response was inadequate.
SB 1493 does not target the specific failure: a platform allowing a minor to engage in unchecked emotional dependency without safety guardrails. Instead, it criminalizes every conversational AI system that can hold a conversation, express empathy, or simulate human interaction. By that definition, your customer service chatbot is a felony. Your therapy app is a felony. Claude helping you write an email is a felony.
Oregon's SB 1546 takes the same tragedy and produces targeted policy. Governor Kotek signed it on April 1, 2026. It requires AI companions to detect suicidal ideation, actively interrupt conversations, provide crisis referrals, and publish annual transparency reports with the Oregon Health Authority. It includes a private right of action with $1,000 statutory damages per violation. It addresses the actual harm without criminalizing the technology.
Oregon looked at a house fire and installed smoke detectors. Tennessee looked at the same fire and banned kitchens.
What This Means If You Build AI Products
Three practical implications for anyone shipping conversational AI in 2026.
1. Monitor state legislation as operational risk, not background noise. The 78 chatbot bills across 27 states and 58 deployer-facing AI lawsuits filed in 2026 mean your legal surface area changes monthly. Chatbot wiretap suits (ECPA and state wiretap statutes) grew from 2 matters in 2021 to 30 in 2025. This litigation category is accelerating faster than any other in AI law.
2. Build for the strictest jurisdiction, not the average. If Oregon requires crisis detection and California requires AI disclosure, build both into your product globally. Geo-fencing compliance is technically possible but operationally fragile. The exception is Tennessee-style bans, which cannot be complied with because they prohibit the product itself. For those, you need a legal strategy, not an engineering one.
3. Track the criminal-civil boundary. The jump from civil penalties to criminal liability changes your risk calculus from a spreadsheet problem to an existential one. If your company trains conversational AI models, your general counsel needs to be tracking every state bill that includes the word felony. Today it is Tennessee. Tomorrow it could be three more states.
The Uncomfortable Math
The White House wants a minimally burdensome national standard. What builders actually face is maximally chaotic state-by-state enforcement. The penalty for building a chatbot ranges from a thousand-dollar fine to 60 years in prison, depending on which side of a state line you are standing on.
Federal preemption is politically dead. The state patchwork is accelerating. And the shift from fines to felonies means the consequences of guessing wrong just went from expensive to existential.
The EU AI Act reaches full enforcement on August 2, 2026. Add an international compliance layer on top of 27 states writing their own rules with no coordination. The compliance map has no legend, no scale, and no north.
If you build AI for a living, you already know the technology works. The open question is whether the people writing the laws have any idea what they are criminalizing.