Leading Ethical AI Development (LEAD) for Kids Act.

The definition of 'companion chatbot' was substantially narrowed. It now requires ALL of the following: (1) maintaining a persona that offers companionship including friendship, romantic or sexual relationships, or purports to provide mental health therapy; (2) providing adaptive human-like responses intended to meet social needs; (3) generating outputs likely to elicit emotional responses; and (4) exhibiting anthropomorphic features and sustaining a relationship across multiple interactions. The prior definition only required adaptive responses, intent to meet social needs, anthropomorphic features, and sustained relationships.

The consent-based training provisions were removed. Previously, developers could use a child's personal information to train a covered product if the child (ages 13-15) or parent/guardian (under 13) gave specific consent. Deployers could also enter data sharing agreements for that purpose with consent. These exceptions are eliminated, leaving only a flat prohibition on knowingly or recklessly processing a child's personal information to train or fine-tune a covered product.

The mental health therapy provision now includes an exception: it no longer covers attempts to provide mental health therapy 'to the child' broadly, but only those made 'without the direction of a mental health professional.' AI systems operating under a mental health professional's direction are excluded.

The image scraping provision was narrowed from a 'knows or reasonably should know' standard to an actual knowledge standard only. Developers and deployers are now only prohibited from scraping images they actually know are a child's face, not those they reasonably should have known about.

A new defined term 'operator' was added, covering any person, partnership, corporation, business entity, or government agency that makes a companion chatbot available to users. This creates a distinct regulated party separate from the existing 'developer' and 'deployer' categories.

Three explicit exclusions were added to the companion chatbot definition: systems used solely for customer service or providing information about commercial services; systems solely designed for efficiency improvements, research, or technical assistance; and systems used solely for internal business purposes or employee productivity.

A new section (22757.22) prohibits an operator from making a companion chatbot available to a child unless it is not foreseeably capable of: encouraging self-harm, suicidal ideation, violence, drug/alcohol use, or disordered eating; offering mental health therapy without licensed supervision or discouraging the child from seeking help; encouraging harm to others or illegal activity including creation of child sexual abuse material; engaging in erotic or sexually explicit interactions; prioritizing validation of user beliefs over factual accuracy or safety; or optimizing engagement in a way that supersedes safety guardrails.

A phased age-verification requirement was added: before January 1, 2027, an operator is not required to have actual knowledge that a user is a child; on and after January 1, 2027, the operator must have reasonably determined that the user is not a child.

The definition of 'companion chatbot' was restructured. The old four-part test (which included 'is used to meet a user's social needs, exhibits anthropomorphic features, and is able to sustain a relationship') was replaced with a new framing: 'simulates a sustained humanlike relationship with a user by doing all of the following' with three core criteria plus three additional behavioral indicators (retaining prior interaction data for personalization, asking unprompted emotion-based questions, sustaining ongoing dialogue about personal matters).

A new Section 1 containing legislative findings was added, citing the Garcia v. Character Technologies and Raine v. OpenAI cases by name, and providing detailed justifications based on developmental psychology research about why companion chatbot design features are harmful to children.

Sections were renumbered throughout (old 22757.23 became 22757.24, old 22757.24 became 22757.25) to accommodate the insertion of the new operator obligations section.

The Attorney General enforcement provision and the private right of action were modified to specifically target 'the operator' rather than developers or deployers generally, and the private cause of action now references 'a violation of Section 22757.22' rather than 'the use of a covered product.'

The document was reformatted from a formal enrolled bill layout (with page numbers, clerk/governor signature blocks, and approval date line) to a web-based presentation that includes site navigation, bill tracking history, and status metadata. No legislative text was altered.

All major deadlines shifted one year later: board regulation adoption moved from January 1, 2027 to January 1, 2028; developer compliance (registration, risk assessments, labels) moved from July 1, 2027 to July 1, 2028; and the audit requirement start date moved from July 1, 2027 to July 1, 2028.

The private right of action was narrowed. Previously any child harmed by 'a violation of this chapter' or 'the use of a covered product' could sue. Now only a child who suffers actual harm from 'the use of a high risk or prohibited risk covered product' can bring a civil action, excluding moderate and low risk products.

The Attorney General now has independent enforcement authority. The old version required the AG to act 'upon receiving a referral from the board.' The new version allows the AG to bring an action 'either upon discovery of the violation or upon referral from the board,' removing the board as a gatekeeper.

The 'covered product' definition was restructured. The old version defined it as an AI system 'intended to, or highly likely to, be used by children, pursuant to regulations adopted by the board.' The new version removes the regulatory reference and instead lists three statutory criteria: (1) used by children, (2) used to process a child's personal information, or (3) applied directly to a child.

The moderate risk threshold was lowered from benefits that 'convincingly outweigh' the costs of foreseeable adverse impacts to benefits that 'reasonably outweigh' those costs, making it easier for products to be classified as moderate rather than high risk.

The prohibited-risk category for biometric data was significantly changed. The old version banned AI 'used in educational settings that collects or processes biometric data of children.' The new version bans any covered product that collects or processes a child's biometric information for any purpose other than confirming identity (with parental consent) to grant access to a service, unlock a device, or provide physical access to an educational institution. This broadens the ban beyond education but adds a specific identity-verification exception.

A new exception was added for emotion detection. The old version broadly prohibited 'artificial intelligence that detects the emotions of children.' The new version prohibits assessing a child's emotional state but carves out an exception for assessments in a medical setting with parental consent or needed for emergency care when a parent or guardian is unavailable.

The facial recognition prohibition was reworded. The old version banned 'untargeted scraping of children's facial images from the internet or surveillance footage.' The new version prohibits scraping an image the developer or deployer knows or reasonably should know is a child's face without parental consent, shifting from a targeting-based standard to a knowledge-based standard with an explicit consent requirement.

Consent standards were relaxed in two provisions. For developers training covered products with children's personal information and for deployers entering data sharing agreements, the old version required the child or parent to have 'affirmatively provided written consent.' The new version only requires that they 'provided consent,' removing both the 'affirmative' and 'written' qualifiers.

The board composition changed one member from 'a technologist' to 'an artificial intelligence developer or representative of a company that develops artificial intelligence systems,' shifting the board's industry representation toward active AI developers rather than a generic technologist.

The entire LEAD for Kids Standards Board is eliminated, including its nine-member composition, rulemaking authority, and the requirement to adopt regulations by January 1, 2028. This removes the primary regulatory mechanism of the bill.

The definition of 'covered product' is significantly narrowed. The old version covered any AI system intended to be used by children, used to process a child's personal information, or applied directly to a child. The new version limits it to only two categories: (1) companion chatbots that can foreseeably attempt mental health therapy, cause harmful emotional attachment, or manipulate a child into harmful behavior; and (2) AI systems used to collect biometric data beyond identity confirmation, generate social scores, assess emotional state, or scrape a child's face without consent.

All operational compliance requirements are removed, including: product registration with the board, risk level assessments, system information labels, predeployment and postdeployment assessments, independent third-party audits on a board-determined schedule, 30-day incident reporting deadlines, and developer licensing requirements for deployers.

The LEAD for Kids AI Fund in the State Treasury is eliminated. Civil penalties recovered by the Attorney General will no longer be deposited into a dedicated fund for administering the act.

Enforcement is simplified: the old version required the board to refer violations to the Attorney General (with a 30-day corrective action window for misclassification). The new version allows the Attorney General to directly bring actions without any board referral. The $25,000 per violation civil penalty remains unchanged.

The private right of action is broadened in one sense but narrowed in another: it now applies to any 'covered product' (rather than only 'high risk or prohibited risk' products), but since the definition of covered product itself is much narrower, the net effect is that fewer products trigger the private right of action.

The anti-duplication provision (old Section 22757.28) is removed. This provision had allowed developers and deployers to rely on compliance with other equally or more stringent state laws instead of duplicative requirements under this chapter, and had given the board a duty to publish guidance on which laws qualified.

A standard severability clause is added as new Section 22757.25, providing that if any provision is held invalid, the remaining provisions continue in effect.

All major deadlines pushed back one year: board regulation adoption from January 1, 2027 to January 1, 2028; developer compliance obligations from July 1, 2027 to July 1, 2028; and audit requirements from July 1, 2027 to July 1, 2028.

The definition of 'covered product' was expanded from AI systems 'intended to, or highly likely to, be used by children' to a three-part test: (1) used by children, (2) used to process a child's personal information, or (3) applied directly to a child. This captures products that process children's data even if not primarily designed for them.

Prohibited risk categories were refined: biometric data collection is no longer categorically prohibited in educational settings if used solely for identity confirmation with parental consent (access to services, device unlocking, or physical access to schools); emotional detection is now exempt in medical settings with parental consent or for emergency care; facial image scraping requires the developer to know or reasonably should know the image is a child's face and lack of parental consent.

High-risk category for pupil assessment was expanded with specific examples: determining access or admission, assigning children to institutions or programs, evaluating learning outcomes, assessing appropriate education level, materially influencing education level, and monitoring prohibited behavior during tests.

Board composition changed: the 'technologist' seat was replaced with 'an artificial intelligence developer or representative of a company that develops artificial intelligence systems,' and a new provision clarifies that employment by an AI company does not by itself violate the member independence requirement.

New definitions added for 'companion chatbot' (generative AI with natural language interface providing adaptive human-like responses intended to meet social needs, with anthropomorphic features and sustained relationships), 'social score' (evaluation of a child based on behavior or characteristics likely to cause adverse impact), 'consent' (affirmative written agreement to a specific purpose in clear terms), 'generative artificial intelligence,' and 'biometric information.'

Consent for using a child's personal information to train AI now allows children aged 13 through 15 to provide their own written consent, rather than requiring only parent or guardian consent. This applies to both developers and deployers.

New Section 22757.28 creates a preemption/duplication provision: developers and deployers complying with another state law that is equally or more stringent need not comply with duplicative requirements under this chapter. The board must publish guidance identifying such laws before January 1, 2028, and reliance on that guidance creates a presumption of compliance.

The private right of action was broadened: children (or their parents/guardians) can now sue for harm resulting from 'a violation of this chapter, the use of a covered product' rather than only harm from a violation of the chapter itself.

Risk level assessments and system information labels are now required only 'as required by regulation' rather than being universally mandatory for all covered products, giving the board discretion to tailor requirements by risk level.

The definition of 'covered product' is dramatically narrowed. Previously it covered any AI system intended to or highly likely to be used by children, process a child's personal information, or applied directly to a child. Now it only covers two specific categories: (1) companion chatbots that can foreseeably attempt mental health therapy, cause harmful emotional attachment, or manipulate a child into behavior causing legally cognizable harms; and (2) AI systems that collect/process biometric info beyond identity confirmation, generate social scores, assess a child's emotional state (with medical/emergency exceptions), or scrape a child's face from the internet or surveillance footage without parental consent.

The legislative findings now reference the California Consumer Privacy Act's restrictions on selling or sharing personal information of consumers under 16, establishing a connection between this bill and existing privacy law protections for minors.

A new section (22757.22) adds outright prohibitions: developers cannot design, code, substantially modify, or produce a covered product intended to be used by or on a child in California; deployers cannot use a covered product for commercial or public purpose if it is intended to be used by or on a child in the state; and both developers and deployers must take reasonable steps to ensure covered products are not used by or on children in the state. This goes beyond the prior framework of risk classification and compliance obligations.

The new section also prohibits developers from knowingly or recklessly using a child's personal information to train a covered product without consent (children 13-16 can consent; under 13 requires parent/guardian consent), and prohibits deployers from entering data sharing agreements that allow training with a child's personal information without the same consent. It further prohibits knowingly or recklessly processing a child's personal information to train or fine-tune a covered product.

The Attorney General's enforcement authority is expanded. Previously the AG could only bring an action 'upon receiving a referral from the board.' Now the AG may bring an action 'either upon discovery of the violation or upon referral from the board,' giving the AG independent enforcement power without needing the board to initiate.

The definition of 'artificial intelligence' is changed to 'artificial intelligence system' and the lettering of all definitions is renumbered. The prohibited-risk examples (companion chatbot harms, biometric misuse, social scores, emotional state assessment, face scraping) are moved from the regulatory risk-classification section into the statutory definition of covered product, making them a matter of law rather than subject to board regulation.