I spent four days at IBC in Amsterdam, 11 to 14 September. First the attendance numbers: According to the closing press release, 41,225 attendees from 170 countries came, with more than 1,300 exhibitors. 256 companies were there for the first time, up 36% on 2025 (IBC). In short: It felt as busy as every year. In today’s issue let us try to better understand what is going on, with a focus on the sub-area of AI and language technology. This issue is mostly about my impressions: stands visited, people talked to.

Only 0.16% of sampled news images carried C2PA content credentials, according to the first run of Metawatch, a new crawl by the International Press Telecommunications Council (IPTC) covering 6,869 images from 441 publishers in 122 countries. It also found that resizing images during publishing strips the credentials, and many publishers do not know it (via Paul Melcher, LinkedIn).

Why care? In the near future any type of content will have to “negotiate” with AI engines. To do this properly and with success, the content industry needs provenance data that is reliable and survives the pipeline. At IBC this year at the Fraunhofer IDMT booth, the discussion pointed to combining several standards, such as C2PA credentials and ISCC fingerprints (International Standard Content Code, ISO 24138) plus custom additions, which might be developed per publisher or per content type.

We don’t need to discuss one thing anymore: AI now works for language, to an extent. What buyers need though is more transparency about where the quality breaks.

Many offers, no way to compare them. First of all there are many offerings in the language tech field. Almost every stand claimed more languages and to provide a full solution. What was missing was benchmarking a buyer can use: which language pairs work well, and how much production time a tool saves. Offering several engines to cover all language pairs is now standard, not a differentiator.

The middle languages are improving. I cannot prove it with hard data, but visitors to the plain X booth kept saying: The “middle” languages are getting better. Examples: Latvian, Estonian and Hungarian were perceived as “quite good” compared to a year ago. The catch is that you cannot see this yourself. You need a native speaker to tell you.

Demand is more specific. The market “learns” about AI-driven language technology platforms, based on observations from Sebastião Miranda, lead developer at Priberam. He said the conversations at the booth were more concrete than last year. People came with particular questions. For example: Interest in audio description came up often, driven by the European Accessibility Act, in force since June 2025.

Live transcription works, but not everywhere. Demand for live transcription and translation is rising. Conferences are the clearest case. In this setting the short delay between the spoken audio and the transcript is enhancing the experience: You hear the speaker, then read what they said a second later, and the text confirms what you heard.
Slight caveat: This is not true when people speak very, very fast (e.g. live sports). Plus: Every system still gets some unfamiliar names wrong, or loses a guest who mumbles. At a conference nobody complains. On air it is a problem for broadcasters.

Subtitles: the fitting is the real work. The industry has accepted the Netflix rules for proper subtitling as the standard: Two lines, a reading-speed limit, timed to the cut. Subtitling experts at the show agreed that the biggest task is not translation but fitting a fast speaker into two lines before the shot changes - especially if the feed is live. Some engines do this by now, for example by shortening a sentence or dropping a word so the subtitle fits.

Mixed languages. At the Speechmatics booth I saw a demo of an engine switching from German to English mid-sentence, with no manual language selection. This matters because much of the world's speech mixes languages: Arabic in francophone regions, for example, carries French words inside Arabic sentences.

Emotion. Panjaya, based in Tel Aviv, synchronises lip movement and body gesture to the dubbed track. I tried it at the booth: ten seconds of English, returned in German (Panjaya).

Phont, still in pre-beta, makes subtitles SHOUT: animated captions whose typography changes with the emotion, with a Premiere Pro plugin, live transcription and an API (Phont). Overused it would be too much, but a stadium "Goaaaaaaaaal" would be a good fit.

Languages & The Media · 4 to 6 November 2026 · Senate House, London · languages-media.com The conference on audiovisual translation and media localisation.

Human-level translation quality is still years away, says Marco Trombetti, CEO of Translated. Speaking at SlatorCon in San Francisco, he argued that the limits of general-purpose AI create an opening for specialised services, and that knowing how to reach top quality is itself the expertise (Slator).

I am Mirko Lorenz. I work on language technology projects at Deutsche Welle in Germany.

Three projects you will hear about in this newsletter:

AI use: Full disclosure, I do use Claude (Anthropic) to research and edit this newsletter, with prompts I have refined many times. My goal is to find out where AI is reliable and where the hallucinations come in. Before publishing I check all facts and links.

Error log: I maintain an open Google Doc where I collect the small and big problems that a tool like Claude introduces to editorial work. The idea is to get a better understanding of where AI is good and where it is not. Read it here. Responsibility for stated facts, names, and links is entirely mine.