Humaina

If you're into #AIregulation, add Turkey to your watch list. 🇹🇷 #Turkey is a unique market, being at the crossroads between Asia, Europe and Middle East. ▪Geographically, it's located in the heart of Eurasia (mostly in Asia minor). ▪Politically, modern Turkey is inspired by European civil law models. ▪Culturally, Turkey has a predominantly Muslim demographic. Given its central position, it's interesting to see which 'side' Turkey will follow on blurry policy issues like AI regulation, which has seen many diverse approaches across the East and West. So far, Turkey is following the #EUAIAct, as announced in its Medium-Term Program 2024-26. On 24 June, Turkey introduced the draft "Artificial Intelligence Bill No. 2/2234". Similar to the EU AI Act, this bill will broadly regulate AI systems based on risk level (e.g. prohibited systems and high-risk systems): https://lnkd.in/gRTfG_ZY Key aspects of the bill include (based on English sources): 1️⃣ Applies to stakeholders like "providers", "distributors", "importers" and "users". 2️⃣ Mandatory principles on development of AI systems (i.e. security, transparency, fairness, accountability, and privacy). 3️⃣ High-risk classification of certain AI systems (e.g. autonomous vehicles, medical diagnostic systems, AI in law enforcement) require further precautions (e.g. a conformity assessment + registration requirements). 4️⃣ Penalties for breach - e.g. TRY 35 million or 7% of annual turnover for deploying prohibited AI systems; TRY 7.5 million or 1.5% of annual turnover for providing false information related to AI. While Turkey is not yet a global leader in #AI, it has a formidable tech sector: ▪Booming innovation hub based on Turkey's young population (median age of 32) and over 100 "technology parks". ▪Strong B2C market, especially in consumer electronics, gaming and e-commerce (with large brands like Vestel and Beko). ▪Well-connected telco infrastructure, with Istanbul being a key hub for underwater cable traffic between Europe and Asia (attracting investment from big cloud companies). ▪Major producer and exporter of unmanned drone tech (though mostly for military use). As Turkey's AI sector continues to grow, I'm curious to see how Turkey will adapt and customise its EU-style AI law over time. In that sense, Turkey fits in a broad category of nations who are leaning towards the EU risk-based model, including Australia, Brazil, Canada, South Korea, Japan and Thailand (though with different levels of alignment). Out of the above countries, Turkey's strong B2C and hardware-focused sector reminds me of South Korea and Japan, both of whom are favouring more lenient "pro-business" regulation. Perhaps Turkey will follow suit? 👓 Want more? Check out my Global AI Regulation Tracker which tracks AI policy and law developments across the world on an interactive world map (see link in the 'Visit my website' button above).

A key feature you cannot forget in your GenAI implementation: AI Guardrails 𝗪𝗵𝗮𝘁 𝗮𝗿𝗲 𝗔𝗜 𝗚𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀? Guardrails are programmable rules that act as safety controls between a user and an LLM or other AI tools. 𝗛𝗼𝘄 𝗗𝗼 𝗚𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀 𝗙𝘂𝗻𝗰𝘁𝗶𝗼𝗻 𝘄𝗶𝘁𝗵 𝗔𝗜 𝗠𝗼𝗱𝗲𝗹𝘀? Guardrails monitor communication in both directions and take actions to ensure the AI model operates within an organization's defined principles. 𝗪𝗵𝗮𝘁 𝗶𝘀 𝘁𝗵𝗲 𝗣𝘂𝗿𝗽𝗼𝘀𝗲 𝗼𝗳 𝗜𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁𝗶𝗻𝗴 𝗚𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀 𝗶𝗻 𝗔𝗜 𝗦𝘆𝘀𝘁𝗲𝗺𝘀? The goal is to control the LLM's output, such as its structure, type, and quality, while validating each response. 𝗪𝗵𝗮𝘁 𝗥𝗶𝘀𝗸𝘀 𝗗𝗼 𝗚𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀 𝗠𝗶𝘁𝗶𝗴𝗮𝘁𝗲 𝗶𝗻 𝗔𝗜 𝗦𝘆𝘀𝘁𝗲𝗺𝘀? Guardrails can help prevent AI models from saying incorrect facts, discussing harmful subjects, or opening security holes. 𝗛𝗼𝘄 𝗗𝗼 𝗚𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀 𝗣𝗿𝗼𝘁𝗲𝗰𝘁 𝗔𝗴𝗮𝗶𝗻𝘀𝘁 𝗧𝗲𝗰𝗵𝗻𝗶𝗰𝗮𝗹 𝗧𝗵𝗿𝗲𝗮𝘁𝘀 𝘁𝗼 𝗔𝗜 𝗦𝘆𝘀𝘁𝗲𝗺𝘀? They can protect against common LLM vulnerabilities, such as jailbreaks and prompt injections. Guardrails support three broad categories of guardrails: 1/ Topical guardrails: Ensure conversations stay focused on a particular topic 2/ Safety guardrails: Ensure interactions with an LLM do not result in misinformation, toxic responses, or inappropriate content 3/ Hallucination detection: Ask another LLM to fact-check the first LLM's answer to detect incorrect facts Which guardrails system do you implement in your AI solutions?

Big news from the NYT this morning -- the AG will lead an antitrust investigation into the AI market with a focus on Microsoft, OpenAI and Nvidia. Although Nvidia may be a surprise to some, it would be difficult to investigate the AI market without also investigating Nvidia, with particular interest on supply of chips. If for example companies have been acquiring large amounts of specialty chips like H100 GPUs to prevent competition rather than for generating revenue, then that could be a prosecutable behavior. It's recently been reported that Microsoft is the largest customer of Nvidia, and Nvidia's CFO said that about 45% of their datacenter business is from Big Techs. Since Big Techs have more money than anyone else, much of which comes from monopolies, if they are using that market power to prevent competition, whether by attempting to buy all the chips, specialty talent, or startups, then that's clearly anticompetitive behavior and probably illegal. There is also no question in my mind and many others that's precisely what's been occurring for many years now. I think it's probably good news that the AG Antitrust Division is leading the investigation as they have a better track record than the FTC. The real question is how much the presidential election year is influencing this. I don't have enough evidence to comment on that other than to say I think the Big Tech antitrust issue is bipartisan, and appears to be becoming a much larger concern across society, industries, and countries.

EVERY SOFTWARE COMPANY seems to be gradually updating its Terms for AI. Here we have an Adobe Photoshop user locked out of the application until he's agreed to the new Terms of Use ✨🎨 One section of the terms reads 👀 "Solely for the purposes of operating or improving the Services and Software, you grant us a non-exclusive, worldwide, royalty-free sublicensable, license, to use, reproduce, publicly display, distribute, modify, create derivative works based on, publicly perform, and translate the Content. For example, we may sublicense our right to the Content to our service providers or to other users to allow the Services and Software to operate with others, such as enabling you to share photos." Users on Twitter (X) have raised the issue that as professionals, they are often under NDA and don't feel comfortable signing up to terms that allow Adobe or others to access their content or train their AI models. "Designer Wetterschneider, who counts DC Comics and Nike among his clients, was one of the graphics pros to object to the terms." "Here it is. If you are a professional, if you are under NDA with your clients, if you are a creative, a lawyer, a doctor or anyone who works with proprietary files – it is time to cancel Adobe, delete all the apps and programs. Adobe can not be trusted." (see comments) Should users be entitled to retrieve your cloud hosted files before having to approve the new terms? Slack recently had to clarify their own terms to their users. Do you expect more software companies to follow this path, in their quest to develop and train their own AI models? Will this encourage more users to self-host files and to seek out open-source applications that don't impose such terms? #ai #adobe #photoshop #terms #tou #nda #llm #genai Source: https://lnkd.in/dvcGyhui