When AI meets Product: May’25 AI Product Updates
Keeping up to date with new AI models, products, ethics, and trends
Welcome to the May edition of “When AI Meets Product — AI Product Updates”. May brought another busy wave of developments in the world of AI, from new foundational model launches to cool applications, important progress around ethics, and enterprise adoption. In this post you’ll find:
- Latest updates from GenAI model providers — Major LLM launches like Claude 4 and updates on models for image, audio, and video generation.
- Interesting new AI applications: everybody is doing general purpose chatbots and trying to monetize them, AI devices are back, and other relevant AI products that will help improve productivity, coding, and learning.
- Updates on the impact of AI to ethics and legislation — relevant resources this month about data and AI privacy, and examples of AI chatbots gone wrong and the ethical problems they can cause.
- AI impact on Product Management and Enterprises — from trends, to how to accelerate adoption and specific company examples.
Let’s get started! 🚀
🧠 Model providers updates
May was packed with updates on foundational LLM versions. The most talked-about release? Anthropic’s Claude 4, with news not only about its improved performance compared to previous versions, but also…
- Internal studies pointing it turned to blackmail when engineers tried to take it offline
- Activation of AI safety level 3 protections (refusing to answer dangerous requests and protecting model weights)
- Some news pointing to a new implementation that would allow the system to contact authorities or press if it detects you’re doing something ‘egregiously immoral’ (with the corresponding risks for business and personal data, user permission, false positives…).
Other relevant LLM launches this month included Mistral’s Medium 3, Amazon Nova Premier , Microsoft’s Phi-4 models, Alibaba’s Qwen 3, and Deepseek R1.
It’s of course a good moment to check out the LLM Chatbot Arena. Just a heads-up though: a recent study raised concerns about its neutrality, suggesting it might favor top labs by letting them test private model versions ahead of others.
Generative AI for images, audio, video, and more also saw exciting updates:
- Google went big, broadcasting many AI news during their annual I/O ’25 (summary video). From the model side, those include Google’s Imagen 4 , Veo 3 and Lyra 2 (image generation model, high quality video generation, and professional-grade audio respectively), Music AI Sandbox (offering musicians a platform to augment their work with AI) and MusicFX DJ (multiple prompts to shape the music in real-time and create an evolving soundscape), and Google Deepmind’s AlphaEvolve (which combines Gemini models with evolutionary approach providing amazing results in algorithm design).
- OpenAI’s latest image generation for API
- Adobe’s Firefly update: professional platform that unifies AI-powered tools for image, video, audio, and vector generation
- Deepgram’s Nova-3 for AI-driven speech to text.
- Seed tars’ UI-TARS-1.5 open-source multimodal agent built upon a powerful vision-language model.
🔥 AI Products, applications & agents
Everybody is doing chatbots…
Everybody is doing general purpose chatbots and testing ways to monetize them (by adding ads, selling products…). #1 application for GenAI might still be AI Chatbots (although some people struggle “what can that help me with?” / “what should I ask it?”), and the number of users for this chatbots continues growing (350M monthly users for Gemini, 400M weekly users for ChatGPT).
- Meta is launching standalone AI app (after integrating Meta AI in most of their products like Whatsapp or Instagram),
- ChatGPT starts offering products and brands, adding shopping to the chatbot experience.
- Googles is expanding ads to AI search and chatbots
All this is impacting how users search and consume information. Google’s AI overviews has been shown to lead to reductions in clickthrough rates from online search results, but at the same time general chatbots are already a thread to Google, as search traffic decreases for the first time in 20 years. Companies will need to adapt their visibility strategies to these new realities!
AI devices are back
AI devices seem to be back! This time by former Apple designer collaborating with OpenAI to create a screen-free device with AI at its core.
Google also presented AndroidXR, augmented reality with LLM integration that displays simlar-to-smart-watches info (texts, time, directions…) but also is fully integrated with Gemini to be able to have a more conversational interaction with the device, and to be able to ask questions where Gemini’s response can use the context of previously seen images or conversations by the user.
Other relevant new AI Products and use cases
Google also announced great AI applications during the last I/O ’25 (summary video). Many of the applications can be mapped to the productivity and personal assistance opportunity:
- Real time speech translation integrated in Google Meet
- Smart reply for emails that are more personalized to your voice
- Integration with different sources of information (notes, mails, pictures) to run smart search and generation on top of them
- AI mode for search allowing for longer and more complex queries, and adding complex analysis and data visualization for sports and financial questions.
- Improved shopping experience (including try on image generation for clothes shopping)
Another area related to productivity is coding assistance to enhance developers productivity. Some relevant applications and updates on that end include Deepwiki (GenAI applied to allow you to understand any repo), Mistral’s agentic LLM for software engineering tasks Devstral, and OpenAI’s Codex updates (AI system that translates natural language to code), and Deep Research connection to Github. On top of this, there was an acquisition showing how hot this topic is: OpenAI acquires Windsurf (for 3bn).
Vibe coding and AI-powered prototyping continues to be a big thing, and this time Figma introduced its own AI vibe coding, and also Google joined the party with https://stitch.withgoogle.com/
There were also other great examples such as:
- Pinterest’s visual search feature
- Google using GenAI for lessons, and expanding LearnLM family of models built with educational experts
- Duoling use of AI to produce new courses
Building Better AI Products: Technical updates
Many companies are preparing for the agentic future. This includes considering MCP (protocol to give agents access to tools), but now another protocol appeared to facilitate collaboration between different agents. The largest payment companies (Visa, Mastercard, Paypal) are racing to get ready to this new world.
All this is boosting in a moment when the technology behind it is not completely reliable. Examples on this is how Bloomberg questions RAG being safer, or potential security vulnerabilities such as Github’s MCP.
⚖️ Ethics & Legislation
There were many relevant resources this month about data and AI privacy. In Common AI Product Privacy Mistakes, some key things to consider are discussed including API-based AI vs local-first AI, the need for evaluations towards privacy concerns, risks of connecting data without understanding origin, and lack of transparency on AI and data use. In another interesting post, more issues related to going for API-based AI were shown: data leakage, compliance violations, opaque behavior…
A more extended AI Privacy Risks & Mitigations Large Language Models (LLMs), covers Data Protection and Privacy Risk Assessment with some examples of LLM Systems’ Risk Assessments that are great to further understand risks and what compliance requirements like DPIA might look in practice.
In terms of legislation, there was a business oriented article on Forbes on how to avoid fines due to EU AI Act (AI system inventory, understand where you are with respect to the four risk categories, and in the case of high risk, deep dive into the new mandatory requirements for thoses systems — governance frameworks, transparency and human oversight, data quality, security…)
What ethical problems can the extended use of AI chatbots carry?
- This month we saw news around Grok AI bringing up “white genocide” in unrelated queries, and holocaust denial. The selection of training data, post-processing and system prompts of these systems can have a huge impact on the type of content they generate, and together with conversational capabilities and access to people’s context, pose a big risk on misinformation and impact on political opinions.
- OpenAI also had a weird case of ChatGPT becoming too sycophantic this month, tending to agree too much with users and even idolizing them. The impact on this on mental health and others is unexplored but Responsible AI principles (accurate, robust, monitored…) could have helped notice this sooner (thoughts on MLOPs techniques that could have prevented OpenAI’s sycophantic responses)
Last but not least, I recently really enjoyed this interview to Ramón L. de Mántaras, with a refreshing and honest view on AI: real capabilities vs what is advertised, problems and incentives to anthropomorphising AI, and much more!
🏢 AI in the Enterprise: How to Get Ready
AI’s impact on Product Management
As discussed in INSPIRED in the Generative AI Era, AI is having two main areas of impact in product teams:
(1) A technology that might help better solve problems
(2) And help product teams discover and deliver products better and faster
With respect to point (1), at Integrating AI for product people you can find an interesting guide on getting started with AI development, from how to recognize and evaluate AI opportunities, to the four data principles for AI success and choosing the right vendor and estimating costs. We have previosly disccussed how Evals are key to develop quality solutions and reduce the risk of embarrassement, and in this cool example of GPTs, you have an LLM eval coach for PMs.
With respect to point (2), accelerated prototyping and coding are kings. In this post you can find an overview of the most important tools in the market for fast prototyping & vibe coding, and a good comparison between them. Beware though, as many who have tried it argue it has big tech debt and security risks!
AI’s impact on Companies
There were also other examples of leading AI adoption and positioning towards an AI first future in different companies:
- A good example are trends to make tech and data closer in the enterprise to be able to stay ahead of the curve in this AI revolution. In the case of trainline this has translated into the CDO taking the CTO role.
- Wikipedia’s approach of augmenting human work instead of replacement
How to get there? Nice example of workshop to drive internal AI adoption. At the same time we have seen some changes in strategy: Klarna rehires humans (after announcing some months back some firing due to work being able to be done by AI — maybe too fast or without accounting for quality?), Duolingo CEO clarified in a LinkedIn post also on the idea of augmentation and not replacement, and overall radiologists as one of the jobs that has always been said will disappear due to AI, are still here.
🛠️ Wrapping it up
That was it from “When AI Meets Product — AI Product Updates”. Another month, another wave of change with AI reshaping products, organizations, and society in real-time. Stay tuned — next month is sure to be just as exciting!
