ALASKA STATE LEGISLATURE  SENATE STATE AFFA.I.RS STANDING COMMITTEE  January 25, 2024 3:33 p.m. MEMBERS PRESENT Senator Scott Kawasaki, CHAIR Senator Matt Claman, Vice CHAIR Senator Jesse Bjorkman Senator Bill Wielechowski Senator Kelly Merrick MEMBERS ABSENT  All members present COMMITTEE CALENDAR  PRESENTATION: CLOUD GENERATIVE A.I. (ARTIFICIAL INTELLIGENCE) OVERVIEW - HEARD PRESENTATION: A.I and DEMOCRACY - HEARD PREVIOUS COMMITTEE ACTION  No previous action to record WITNESS REGISTER  ADDIE COOKE, A.I. Policy Lead Google Cloud Arlington, Virginia POSITION STATEMENT: Presented on Cloud Generative A.I.   ILANA BELLER, Field Manager Public Citizen Richmond, Virginia POSITION STATEMENT: Presented on A.I. and Democracy.  ACTION NARRATIVE 3:33:08 PM CHAIR SCOTT KAWASAKI called the Senate State Affairs Standing Committee meeting to order at 3:33 p.m. Present at the call to order were Senators Bjorkman, Wielechowski, Merrick, Claman, and Chair Kawasaki. ^Presentation: Cloud Generative A.I. (Artificial Intelligence) Overview  PRESENTATION: CLOUD GENERATIVE A.I. (ARTIFICIAL INTELLIGENCE) OVERVIEW    3:34:07 PM CHAIR KAWASAKI announced the consideration of a presentation on Cloud Generative A.I. 3:34:36 PM CHAIR KAWASAKI introduced Ms. Cooke. 3:35:47 PM ADDIE COOKE, A.I. Policy Lead, Google Cloud, Arlington, Virginia, provided an overview of A.I. technology. 3:36:59 PM MS. COOKE moved to slide 2 and stated that A.I. is a transformational technology and Google treats it as a serious revolutionary tool, similar to the transformational changes of steam power, electricity, smart phones, and internet, and A.I. will also prove as a significant shift in our lives. Governments have the tools but need to determine the ways to put them to use. 3:38:34 PM MS. COOKE moved to slide 4 and said that A.I. is not a new technology, but it has become more powerful. In 2017, Google put out a seminal paper on the Transformer, the "T" in ChatGTP. Google Cloud open-sourced its research to allow the entire world to benefit from learnings. Google itself benefited from Bidirectional Encoder Representations from Transformers (BERT), a transformer that supports the Google search function and has allowed users to have information readily at the fingertips. A significant amount of progress has been made since. Ms. Cooke said an open-source model, AlphaFold, was released in 2021 which looks at different proteins to develop new medicines. She recollected an article that unveiled the discovery of thousands of new formulas that address various challenges. These models are in the same vein as Google's search model. The goal is to find patterns across volumes of information that were previously unavailable. However, serious considerations must be made when building revolutionary technologies. In 2017, Google established responsible A.I. principles. She said that as a seat holder on one of the committees for responsible A.I. The committee and the legal team have the opportunity to review every product for regulatory concerns, alignment, and risks. 3:41:17 PM MS. COOK continued that the committee offers support for the documentation on products before the products are made available to customers. 3:41:35 PM MS. COOKE moved to slide 6 on large language models. She said she used old car manuals to illustrate the capability of large language models. By entering the information from the car manuals into a language model, an individual would be able to ask a question on how to fix a car and receive an answer. The model could also identify patterns across different cars. The power of current A.I. is that coders aren't required, so an individual only needs to be able to read and write to access the technology. 3:43:16 PM MS. COOKE moved to slide 7 and said that consumer-facing products at Google have different considerations regarding risk management than the models for enterprises and government. A consumer-facing model is constantly retraining on the data it is fed, which is why Google has prioritized making an enterprise version available. It will provide more security for enterprises and governments. 3:45:09 PM MS. COOKE continued to slide 8 titled "Google Cloud A.I. Portfolio." She said Google Cloud added several foundation models as well as a developer suite of tools. The Enterprise Search model would have been useful in bill drafting during her legislative role during her time in the Texas Legislature. This technology mitigates "A.I. hallucinations," can summarize information, and is more efficient. The duet suite of products includes tools that support a variety of development. 3:48:34 PM MS. COOKE noted that Google takes data very seriously. Any data used by the Google Cloud Tenant is not retrained on user data contrary to some competitors' structure. 3:49:44 PM MS. COOKE moved to slide 10 on GovAI tools. As a former legislative correspondent, she acknowledged the arduous process of drafting legislative. New technology could help facilitate constituent responses and assist in communication efforts. It could quickly identify patterns and produce a complete email using pattern identification. MIT and Stanford performed a study observing call center agents and found that an entry level call center agent using an A.I. powered support engine was able to triage 20 percent more questions. A.I. technology can transform an entry-level employee to work more efficiently and communicate quickly, creating expertise on demand. 3:52:21 PM MS. COOKE moved to slide 11. Using Google Search, results are at times unexpected, so it's important to ensure that the type of A.I. used is fit for the purpose intended. This is common sense risk management, but Google encourages enterprises and governments to consider intent and read the terms of contract as a default, ensuring employees are not retraining the data or improving its performance on an existing task by using a different dataset. 3:54:09 PM MS. COOKE said that Google encourages users to consider cost controls in the decision-making process. 3:54:20 PM MS. COOKE moved to slide 12 with examples of GovAI used today. She said that it is important to determine when A.I. is appropriate to addressing challenges. Some examples include: Camera installation on emergency vehicles to anticipate pothole triaging and assist homeowners in obtaining insurance. She said that there are opportunities to use this technology as long as risks are managed along the way. 3:56:02 PM MS. COOKE moved to slide 13 and spoke to five areas of risk management outlining the following values: privacy, choice/value, factuality, ease of use, and security at scale. 3:56:39 PM MS. COOKE moved to slide 16 regarding Google's A.I. principles. She said that if technology is not in line with principles, Google takes the appropriate steps to ensure it can be further developed before releasing it to the public. Premature products may be held to ensure responsible A.I., such as facial recognition technology, which required additional time prior to release. Google Cloud offers an image recognition system but does not perform facial recognition. It is up to technology users to use it responsibly. It is a shared responsibility. 3:58:24 PM MS. COOKE moved to slide 17. She highlighted line 7 relaying principles A.I. should be made available for uses that accord with the following principles: • Primary purpose and use • Nature and uniqueness • Scale • Nature of Google's Involvement 3:59:09 PM MS. COOKE moved to slide 18 showcasing risk management methods. She said that Google Cloud has risks that are different from other platforms such as YouTube, Google Search, or Pixel. While each company follows the same A.I. principles, there are disparate risk management committees. Department of Labor would have different considerations than the Department of Health and Human Services. 4:00:24 PM MS. COOKE moved to slide 19 and said that data needs to be assessed to ensure it is valid for the model intended and obtained with permission. 4:01:02 PM MS. COOKE moved to slide 20 showing the lifecycle of the A.I. responsibility model: • Define problem • Collect and prepare data • Train model • Evaluate the model • Integrate and monitor 4:01:34 PM MS. COOKE moved to slide 22 and highlighted risk management takeaways. It is necessary to use judgement-based decisions, consider how these technologies might cause harm, and understand transparency requirements, especially in government when interacting with citizens. 4:02:45 PM MS. COOKE moved to slide 23 and said that use within government agencies begins with an idea. A.I. can create, summarize, discover, and automate, and is applicable through government agencies. She offered examples of government use cases. 4:03:21 PM MS. COOKE moved to slide 24 and said that the graphic that lists various considerations for assessing A.I. projects. 4:03:31 PM MS. COOKE moved to slide 25 and said the following is a neat list of organizational readiness considerations. Organizational Readiness Considerations  • Risk Assessment / Align with organizational strategy • Project Governance • Persona-based training and skills development • Architecture governance • Policies and procedures • Software Development Lifecycle (SDLC) Integration • Testing (internal and external) • Reporting (internal and external) • Regulatory watch • Incident response She said that enterprises need to understand the regulatory landscape when purchasing a new technology. Testing and reporting are core to ensuring readiness of A.I. before it is adopted. 4:04:02 PM MS. COOKE moved to slide 27 and said that cross-functional teamwork is needed in government to properly manage A.I. 4:05:02 PM SENATOR WIELECHOWSKI stated that policymakers are wrestling with regulations and statutes that may be needed to address A.I. concerns. Eric Schmidt, former Google CEO, said that A.I. posed existential risk and could cause people to be harmed or killed. Elon Musk warned of the same thing. Jeffery Hinton, the "Godfather of A.I.," warned of A.I. dangers. He asked if legislation should be passed to address these concerns. 4:05:55 PM MS. COOKE replied that key components lacking in A.I. development are security benchmarks and testing. Google has started working with the government to undergo testing to ensure all proper controls are in place to prevent catastrophic harm. Safety filters are considered in development phases. Currently, child sexual abuse material (CSAM), terrorist content, and harmful Chatbot language are filtered out. She noted that filters can be circumvented, therefore, governments need to adopt globally consolidated standards. Working with the government and third parties to understand risks is important to understanding how the technology works and how it can cause risks. ^Presentation: A.I. and Democracy PRESENTATION: A.I. AND DEMOCRACY  4:08:06 PM CHAIR KAWASAKI announced the consideration of a presentation on A.I. and Democracy. 4:09:30 PM ILANA BELLER, Field Manager, Public Citizen, Richmond, Virginia, presented on A.I. and Democracy. She said that she leads State Artificial Intelligence for Public Citizen. Deepfake is fabricated or fraudulent content, such as video, audio, or images depicting a falsification of another individual's actions. A video example was provided on slide 2 depicting Joe Biden speaking about the film "We Bought a Zoo." 4:12:14 PM MS. BELLER moved to slide 3 and suggested that the application of deepfake can be incredibly serious. A couple months ago, a major election was held in Slovakia involving a deepfake recording of one opposing candidate. Within 48 hours of the election, a deepfake recording went viral accusing the other candidate of wanting to rig the election, but the targeted candidate did not have a chance to dispel the falsified statement. It was assumed to have influenced the outcome of the election. 4:13:31 PM MS. BELLER moved to slide 4 and said that deepfakes released by candidates were developed to uplift their own attributes and attack their opponent. This has been seen in the U.S. A recent robocall with an A.I. voice resembling President Joe Biden targeted thousands of New Hampshire voters urging residents not to vote. Experts say, "2024 is going to be the first true A.I. election in the U.S." She said the two other reasons regulations are urgently needed are: 1. Deepfake technology is becoming rapidly more accessible. Anyone can easily create a deepfake in two minutes. 2. Deepfake technology is improving rapidly in quality. She stated that in August of 2023, the National Institutes of Health (NIH) implemented a study which found that 27 to 50 percent of people could not identify a deepfake. Technologists have said that they soon may have difficulty deciphering reality from a deepfake. 4:16:30 PM MS. BELLER said that these technologies are rapidly more accessible. It is crucial these technologies are regulated as soon as possible. It is not unrealistic that a deepfake could swing the outcome of the U.S. election. Someone could potentially defraud and manipulate voters into voting a certain way. She opined that election influence is not only the immediate concern. There are also larger societal concerns that need to be considered, involving deepfakes, which contribute to the erosion of social trust. An increase in deepfake content will make it challenging for anyone to know what to trust. A bad actor politician could falsely blame their own wrongdoings caught on camera on deepfake fabrication. This technology creates a whole ecosystem of disinformation and everything becomes questionable. Society is at this point to some degree, so further eroding trust in our democracy should be mitigated. 4:18:49 PM MS. BELLER moved to slide 5 and spoke to legislative solutions and referenced five states that have passed legislation to address deepfake concerns. Twenty-six states have introduced legislation and 10-11 other states are actively working on drafting legislation. 4:19:52 PM MS. BELLER moved to slide 6 and relayed that Public Citizen has implemented a tracker that tracks passed or introduced legislation related to deepfakes. She emphasized that there is no discernable patter to the groups or individuals advocating for the legislation. 4:21:19 PM MS. BELLER moved to slide 7 and spoke to key elements of anti- fraudulent deepfake legislation. [Original punctuation provided.] Key Elements of Anti-Fraudulent Deepfake Legislation  • Prohibit distribution of unlabeled deepfakes within [90] days of election o Why disclosure? • Standard: "Deceptive and fraudulent" = Shows a person saying or doing something that they did not say or do. • Cover all persons not just candidates, parties and committees • Establish the requirement for disclosure as prominent as other text, spoken plainly, etc. • Establish a right for affected parties to seek injunction to take down. • Establish enforcement and penalties Ms. Beller added that "disclosure" legislation is necessary when considering first amendment concerns. One of the biggest risks may arise from social media influencers, who might knowingly put out fraudulent deepfake content to defraud millions of people. 4:24:48 PM MS. BELLER moved to slide 8. [Original punctuation provided.] Protections  • No liability for broadcasters or platforms that make reasonable effort to prevent deepfakes, or that show deepfakes as part of news coverage and describe as deepfakes • Exception for satire • Severability 4:25:46 PM MS. BELLER concluded that Public Citizen is working with several legislators to look toward regulations on deepfake. While it is a real, dangerous issue, there is a real solution to addressing it. Forty-one states are currently in the process of taking legislative action. 4:26:50 PM SENATOR CLAMAN asked why the push for legislation is on the state rather than federal level. He suggested that it sounds like interstate commerce. 4:27:29 PM MS. BELLER said it would be great if this legislation was passed on a federal level. Public Citizen supported a bipartisan group of congresspeople, involving Senator Klobuchar and Senator Hawley, to draft a bill, but it is unlikely to pass ahead of the national election. The legislation would only apply to federal election candidates rather than on the state level. Public Citizen also put forth a petition, but the U.S. Securities and Exchange Commission (SEC) has been slow to take action, so it is unlikely laws would be established ahead of the election. States play an important role regardless of whether federal legislation is passed. 4:29:12 PM SENATOR CLAMAN asked whether individual candidates currently have to bring a cause of action to civil court to stop fraudulent action or if legislation is needed. 4:29:32 PM MS. BELLER said legislation is needed because there is no current legislation stating that putting forth a deepfake is illegal. 4:29:46 PM SENATOR WIELECHOWSKI asked Ms. Cooke if technology exists to identify whether videos are real. MS. COOKE replied that Google has a watermarking and verification technology in development called synth ID (synthetic identification). 4:31:17 PM MS. COOKE continued that it is important to continue having conversations about deepfake. The approach must be standardized on a transnational and global level. Resources put behind solving the problem are resources well spent. There will always be a solution, but it requires partnerships and a consensus across the globe. 4:32:32 PM SENATOR WIELECHOWSKI asked if the technology is advanced enough to solve election concerns or if consideration has been given to legislation requiring synthetic identification filtering. 4:32:59 PM MS. COOKE said until an agreement is reached, it is difficult to create a mandate, but a market for these technologies remains. Even the best technology can be subverted by sophisticated actors, so educating the public on fact checking is crucial. 4:34:11 PM MS. BELLER said Public Citizen supports watermarking, but it is not currently possible to prove provenance. To clarify, deepfake legislation focuses on the circulator rather than the creator. 4:35:22 PM CHAIR KAWASAKI said that the legislature considered passing A.I.-focused legislation last December. However, there were challenges with finding a balance to recognize both the positive and negative attributes of A.I. technology. 4:36:11 PM MS. COOKE replied that in the context of security, A.I. must be fought using A.I. One solution on the market is a security large-language model. Google Search receives an enormous volume of hacking attempts, therefore the company trained a model to identify threats entering the system using 20 years of training. While users weren't well educated on social media in its initial release, there is now an opportunity to educate the public on A.I. technology, "A.I. hallucinations," privacy protection, etc. When working with an enterprise customer, A.I. is used to scan data to ensure personally identifiable information (PII) has been removed before data is accepted. 4:39:33 PM SENATOR BJORKMAN asked about communications labeling and first amendment concerns surrounding political advertising. He wondered if there are additional first amendment concerns that would prevent a law from standing a court challenge. 4:40:21 PM MS. COOKE replied that having an exemption for satire is important for first amendment rescinds, but Google has found no further first amendment concerns preventing legislation from moving forward. 4:41:10 PM CHAIR KAWASAKI referenced the first deepfake model that was broadcasted in New Hampshire. He asked what in this case would prevent someone from circumventing loopholes under current laws. 4:42:01 PM MS. COOKE replied that the case would have to play out in court. She suggested that there may be plausible deniability for audio- specific deepfake content. 4:43:11 PM CHAIR KAWASAKI said that the current legislative budget allots DPS and other agencies a large sum of money to cover the legal cost of redacting personal information from public information for the press. He asked for more information about the redactions of Personal Identifiable Information (PII) using A.I. 4:43:57 PM MS. COOKE clarified that the question is about any technical tools available to redact state-level FOIA requests to expedite responses to constituents. She offered to follow up on this use case to find solutions. 4:44:45 PM SENATOR MERRICK opined that the presentation was interesting. She asked if law enforcement agencies have expressed concerns about the fabrication of evidence. 4:45:09 PM MS. COOKE responded that it has not been expressed as a concern. 4:45:12 PM CHAIR KAWASAKI thanked the testifiers. The Legislature has been thinking about ways to navigate A.I., and Senator Shelly Hughes recently introduced legislation on political deepfakes. 4:47:38 PM There being no further business to come before the committee, Chair Kawasaki adjourned the Senate State Affairs Standing Committee meeting at 4:47 p.m.