10 important ideas from TED AI Vienna 2025

One of my career highlights was getting invited as a panelist for TED AI Vienna. But the bigger surprise was how cool this event was. ❤️‍🔥 The line-up was absolutely stellar (see for yourself here) and the talks were not just inspiring, but also intensively intellectually stimulating. 👌🧠

I left smarter. And I came back in Bucharest to share some of my freshly remixed ideas at Unfinished Festival (yesterday) and How to Web (this Thursday).

so even though my brain is a bit over-solicited these days, I was able to squeezed my messy memory into 10 ideas I think we should all reflect on. If you like thinking about tech with a little poetry and a lot of practicality / read on 💋

⚠️ but, before we get started: TED AI is all about Ideas Worth Spreading and I would really appreciate if you would share this article if you find it meaningful 💌 it helps me grow my little newsletter world 🌱 

now here are some of the best ideas I left with, in no particular order (they’re numbered for ease of reading) 🧠

1. There’s a great AI divide

2.6 billion people lack internet access, creating a massive inequality gap. AI is developing faster than any previous technology, raising the critical question: will it close inequality gaps or dramatically worsen them by amplifying existing disadvantages?

I’ve been raising it in my talks as well, the topic of the deepening gap and the further empowering of the powerful, but it was interesting (and painful?) to see the numbers.

👉 source: Bolor-Erdene Battsengel

2. Superhuman persuasion is already here

Current AI is more persuasive than 82% of humans and 21% more effective at changing political opinions. With access to personal data, this gap widens to 82% - representing an immediate threat that requires urgent regulation before superintelligence even arrives. Should all bots be “forced” to reveal they’re not human?

👉 source: Philipp Kloeckner

3. AI is making us intellectually lazy

(I was hoping it’s not true, but) Knowledge workers are becoming "intellectual tourists" in their own work. Research shows AI-assisted groups produce fewer ideas collectively, workers think less. Also, it seems that people remember less when relying on AI for writing / summaries.

And there’s a problem at the level of metacognition: we are slowly losing ability to think about our thinking process.

Advait Sarkar (Researcher @ Microsoft, Lecturer at Cambridge + UCL, and my co-panelist at TED AI Vienna) ended his brilliant talk with this line:

👉 source: Advait Sarkar

5. We’re witnessing the death of traditional scientific research

Enter AI Scientists 🦾🧪 AI is solving 50-year problems like protein folding in minutes and achieving Nobel Prize-level breakthroughs. We're moving toward fully autonomous AI scientists that design experiments, analyze data, and generate new hypotheses - potentially replacing the traditional PhD research process.

Machine intelligence is slowly becoming the ultimate scientific tool. Scientists’ role shifts to asking the right questions (I think similar to the design role 🙃)

👉 source: Oriol Vinyals

6. AI can help turn human trust into an algorithmic currency

In a brilliant talk that proved that tech products can have REAL social impact, Mercedes Bidart showed us how her work is using AI to detect financial reliability from text messages, one-minute videos, and social media posts with 83% accuracy. This enables credit access for 99% of Latin American micro-businesses that lack formal paperwork, transforming informal economies into formal financial systems.

In this way, AI becomes a tool for good, that helps people “see value where others see risk” and “ gold where others saw stones”. It also becomes an instrument of hyper-personalization of financial services honouring local knowledge and culture, by focusing on “milions of quiet signals” showing reliability rather than traditional bank statements.

People stood up to clap for this talk and I have a huge crush on Mercedes 🩷

👉 source: Mercedes Bidart

7. Sousveillance: Watching the Watchers

Pau’s stellar talk explored the concept of a Reversed Panopticum powered by AI, where citizens could us AI to monitor those in power. Examples include tracking parliamentary sleepiness and police accountability, creating bidirectional accountability systems.

I love this counter-narrative to corporate and government surveillance, through "surveillance from below" and Pau’s delivery was exceptional (and he was my conference bestie).

He made an important point: technology creates power asymmetries: between society and government, between society and corporations - and AI gives us the opportunity to rebalance power and create bidirectional accountability systems.

You should check all the amazing work Pau and his team are doing at Domestic Data Streamers.

👉 source: Pau Aleikum Garcia

8. Shadow AI in organizations is real

95% of organizations fail at AI implementation, but individuals achieve 40% productivity gains through "shadow AI" use. This creates adversarial cycles where companies penalize AI use while employees hide their usage through anti-detection methods.

This idea felt so unnatural and ineffective. Some solutions were also suggested: normalizing AI use through transparency (Intercom’s co-founder Des Traynor offered their internal example of labeling ChatGPT-generated docs), making AI use explicitly safe within companies policies, focus on participatory design involving impacted users.

👉 source: a panel titled: What needs to be done to successfully roll-out agentic frameworks across organisations and how can they maximise human potential?

9. AI systems are creating a learning efficiency revolution

More powerful AI models paradoxically need less data to learn. Reasoning models can prove mathematical theorems from single papers and learn strategies from minimal examples, breaking the traditional requirement for massive datasets.

Historical progression demonstrates this principle:

• RNNs: First powerful learnable models (could translate without overfitting)

• Transformers: More computational power + maintained learnability

• Reasoning models: Current frontier with thinking capabilities

So basically, reasoning models are changing the rules: rather than requiring massive datasets, they can learn strategies from a small number of examples and dynamically retrieve and process information instead of memorizing facts. Currently, these models still rely on verified training data, large pre-trained foundations, and predominantly sequential reasoning that limits parallelism. The next advance will be models capable of learning from arbitrary sources (including non-structured material, like poems 🙃), performing parallel reasoning, and acting as research-grade agents able to support scientific discovery. That future may be nearer than expected but timelines remain uncertain, and challenges around data verification, scalability, and dependency on foundation models must be resolved first.

👉 source: Lukasz Kaiser

10. AI can be used as an Ecological Detective

AI is being used to search for endangered plant species across vast forest areas using drone photography and synthetic data generation. This represents a new form of technology-assisted environmental activism addressing the "silent extinction crisis".

👉 source: Laura Cinti

whoa, we’ve been through a lot together! 🧠🫀

and this only scratches the surfaces of all the intellectual magic I experienced at TED AI Vienna this year.

💬 Here’s a challenge for you: comment with (or DM me anywhere) the idea(s) that would want me to expand on in future newsletter issues.

in the end, I’ll leave you with a nice shot from my TED AI panel + the invitation to learn more about AI x Design by taking my course on Interaction Design Foundation (this link includes a 25% discount).

⚠️ Make sure you get my future newsletters in your inbox by subscribing here:

hugs hugs

Ioana 🪩

/ soft machines