The BBC is leaning into imperfect camera footage for Blue Planet III to prove its wildlife scenes are not computer generated. Elsewhere, Jack Dorsey's Block has unveiled an expressive on-device voice assistant, and a growing list of once-hyped AI projects is shutting down.
In this round-up
The BBC Natural History Unit has decided to keep flawed footage in its upcoming series Blue Planet III rather than cutting around it. In the past, production crews would discard shots featuring floating plankton, murky water, or man-made clutter such as buoys and boats. However, because generative AI video often looks hyper-polished and unnaturally clean, Executive Producer Elizabeth White told The Guardian that the production team is intentionally leaving in the rough edges to reassure viewers that what they are seeing actually happened.
For live events and video production teams, this highlights an interesting shift in audience expectations. High-end delegate video, highlight reels, and sizzle clips have spent years chasing spotless perfection. With synthetic media now flooding social channels, polished gloss can make real work look fake. In corporate video and live coverage, leaving in authentic, spontaneous moments, background movement, and natural audio will increasingly serve as your best proof of authenticity. There is nothing to install here, but it is worth remembering when you brief your next videographer.
Read more: theverge.com
Tech company Block, headed by Jack Dorsey, has revealed a research model called Dessa that handles natural speech without converting it into text first. Most voice tools translate your words into text, run it through an underlying large language model (the technology behind ChatGPT), and turn that back into audio. Dessa instead trains directly on human speech. The resulting voice includes sighs, pauses, laughs, and emotional breaks, and it runs locally on consumer computers such as an Apple MacBook M4 Max.
Block says the system is full-duplex, meaning it talks and listens simultaneously rather than waiting for an awkward pause. There is no public release yet, as Block is running safety evaluations to prevent conversational scams. When this style of voice tech arrives in consumer products, it could replace the stiff, robotic automated helplines used for delegate enquiries or venue ticketing with something that feels remarkably natural. For now, it is an early research preview with no immediate action required.
Read more: techcrunch.com
A new post-mortem of failed tech projects reveals that roughly 42% of corporate AI projects are abandoned, according to data from S&P Global Market Intelligence. Among recent casualties, workflow automation tool Relay shut down in August 2026 after native tools inside larger ecosystems swallowed its features, while popular chatbot app Character.ai filed for bankruptcy.
Even OpenAI has had to trim back experiments. Its standalone web browser ChatGPT Atlas was discontinued after less than a year, with features folded back into the main chat app, while video generation platform Sora was shut down in March 2026 amid steep compute costs and poor user retention. A planned super-app interface redesign was also rolled back following user complaints.
For event agencies and venue operators, the lesson is caution around procurement. Avoid building critical operational workflows around tiny, standalone AI tools that sit on top of third-party platforms. If a tool simply wraps an existing model, the foundation company may build the same feature natively, leaving your team stranded. Stick to established platforms or ensure your event data exports easily.
Read more: techcrunch.com
A study conducted by conversational AI research studio Convovo and Cornell University revealed that major AI models show strong political bias when forced to make electoral recommendations. When researchers queried the developer APIs (the behind-the-scenes connectors software uses to talk to AI systems) to pick a winner in US midterm races, SpaceXAI's Grok chose the Republican candidate in 92% of cases, while OpenAI's GPT model picked the Democrat in 97%. The two agreed less than 10% of the time.
OpenAI noted that its consumer-facing ChatGPT interface refuses to answer such prompts, arguing the test forced unnatural behaviour through developer tools. However, researchers noted that the sources each tool relied upon behind the scenes differed sharply, with Grok leaning on social media platform X and conservative guides, while GPT drew on outlets like the Associated Press, Wikipedia, and Reddit.
If your agency uses AI tools to research seminar panel topics, summarize keynote debates, or brief speakers on public policy, be careful. The model you choose will shape the framing of your research. Never treat AI summaries as neutral background reading for contentious event panels.
Read more: techcrunch.com
Pennsylvania Governor Josh Shapiro has signed the Artificial Intelligence Deceptive Advertising Act into law, making Pennsylvania one of the first US states to mandate clear disclosures on AI-assisted marketing. Under the legislation, any advertisement using text, images, audio, or video that has been significantly modified or created using artificial intelligence must feature an explicit disclosure in the same format as the ad itself.
The rule applies to everything from multi-million-pound television commercials down to local flyers, regardless of where the material was originally designed, provided it runs in Pennsylvania. The law contains broad carve-outs for parody, satire, and content that is evidently fictional, creating questions about how enforcement will work in practice.
While this is an American state law, event organisers who market global exhibitions or run paid digital ad campaigns targeted at US delegates need to keep track of this trend. If you use generated imagery on flyers or social ads, transparency requirements are tightening worldwide. Clearly flagging synthetic assets is quickly becoming standard commercial compliance.
Read more: theverge.com
Personal assistant bots are increasingly ditching custom phone apps in favour of running entirely through standard SMS, Apple iMessage, and RCS text messaging. New tools including Ask Proxy, Caddy, and Comma are designed to live directly inside your phone's default messaging app. Users can text them requests to check calendars, set reminders, schedule appointments, or summarise email threads.
Ask Proxy charges £0.77 ($1) per day for personal users, or roughly £11.50 ($15) to £17.65 ($23) per employee monthly for businesses, aiming to anticipate tasks before you ask. Meanwhile, open-source tool Comma connects across Signal and Telegram to draft briefings and manage to-do lists.
For on-site event managers juggling supplier calls, deliveries, and run sheets, having an assistant that lives inside standard text messaging rather than requiring yet another dashboard login could be genuinely practical. If you run small freelance production teams, testing an SMS-based bot could help keep track of build schedules without forcing crew members to download unfamiliar apps.
Read more: techcrunch.com
A high-profile security vulnerability in the programming language PHP, dubbed PHP2Shell, highlighted how AI is altering cybersecurity response times. Security researchers used advanced AI tools to rediscover and build workable exploits for bugs that had remained dormant in software code for years. While the PHP Foundation confirmed no live systems had been successfully breached, malicious actors launched millions of automated exploit attempts within hours of the public disclosure.
The vast majority of the world's websites, ticketing portals, and venue booking platforms run on PHP. While cybersecurity experts point out that AI models still require skilled human direction to uncover real exploits, the speed at which automated bots can strike once a flaw is known has increased dramatically.
For conference organisers and venue teams managing attendee data and registration portals, this is a clear reminder to speak to your web hosting partners. Ensure your external registration platforms and WordPress sites apply security patches immediately, as automated attacks now begin within hours of flaws going public.
Read more: wired.com