Thibault Schrepel
@profschrepel
Associate Prof VU Amsterdam • Faculty Affiliate Stanford • Into Running 🏃🏻 #antitrust #AI #complexityscience #digitalmarkets 📕 www.thibaultschrepel.com 📻
EU merger law tells companies they can defend a deal with efficiencies. I read all 57 cases where they tried. Accepted as decisive: zero. The defence has never once saved a merger the European Commission wanted to block. It is a promise the record does not keep. Paper coming.
This paper is an important reminder that “open weights” do not automatically create trust. If model trainers can, in principle, plant statistically undetectable backdoors, then relying on open-source frontier models developed by geopolitical rivals becomes a much harder governance question.
The EU already has digital sovereignty. Just the wrong kind. Across 34 laws, five types written into the rules: ➝ infrastructure ➝ adequacy ➝ jurisdiction ➝ access rules ➝ credentials & coin All of it controls foreign tech. None builds any. Audit 👇 thecompetitionsystem.substack.com/p/the-paperw...
I ran the European Commission’s new draft Merger Guidelines against all the merger decisions ever published. In many regards, the Guidelines go against the case law while pretending otherwise. Paper to follow...
Brussels calls the Tech Sovereignty Package the moment Europe gets serious about digital independence. I mapped 34 EU digital laws. Europe has legislated sovereignty since 2014. The legal stack exists. It governs foreign tech, builds none. thecompetitionsystem.substack.com/p/the-paperw...
Summer = research & running ☀️ Living in Amsterdam for eight years is a gift. Why travel elsewhere in the summer? With the canals, forests, the swimming, we have so much right here. This city is so incredibly hard to beat during the summer time.
Two AI agents. 400 rounds of pricing. No agreement, no messages, and yet the price keeps climbing. One of them even suggests they avoid raising red flags. Tanise Brandão and Carlos Neves (CADE) join Alba Ribera Martínez and me on a new Stanford #ComputationalAntitrust episode.
The Google Android fine was never the story. The story is one sentence that retires the as-efficient competitor test in digital markets. It will be quoted for years. My full audit of C-738/22 P 👇 thecompetitionsystem.substack.com/p/the-quiet-...
I am launching a newsletter. It is called The Competition System. It is an empirical, referenced account of the cases and rules that shape competition in Europe. Articles 101 and 102, the DMA, DSA, the AI Act, MiCA, the Data Act, etc. Every claim is traceable to its source.
Wrote a strong Master’s thesis on competition law or competitiveness this year? We are opening the 2nd edition of our award, with ICLE and VU Amsterdam. Three winners present in Amsterdam, travel covered. Deadline 31 August 2026. alti.amsterdam/call-for-sub...
1 – Today’s Google Android judgment (C-738/22 P) is the end of a 13-year case. It’s also the start of something bigger. The EU Court of Justice just confirmed the 3rd generation of big tech competition enforcement. And the Commission is already applying it to AI, before the ink is dry. A thread 🧵
Data neutrality does intensify downstream competition, but it kills the platform's incentive to refine and produce the very input the rule is trying to equalize. The DMA’s data-sharing mandate may be equalizing access to a shrinking pie. Fair ≠ welfare-enhancing.
New #ScalingTheory episode out now, with Steven Pinker. Here, Steven questions whether AI could bring us back together, after social media pulled us apart. ▶️ www.youtube.com/watch?v=ijNj... 🎧 open.spotify.com/episode/36hX... 🍎 podcasts.apple.com/fr/podcast/s...
NEW in Stanford #ComputationalAntitrust (Vol. 6). Piero Malca Vilchez, César Quiñones Costa, and Enzo Gomez Rojas build Ulysses, an LLM workflow that predicts whether conduct is anticompetitive under Peruvian competition law. On their pilot benchmark it scored 11/11, ahead of GPT-5 and Claude 4.1.
LLMs don't have personalities. Those tidy "psychological profiles" people keep extracting from GPT and friends? Largely a measurement artifact. Change how you ask, and the personality changes with it. You're not profiling the model. You're profiling your own questionnaire.
In 1865, Britain classified the car as a “road locomotive.” Speed capped at 4 mph. A man with a red flag had to walk ahead of it. Britain invented the engine. France and Germany built the car industry. New paper: the DMA is about to do this to AI. papers.ssrn.com/sol3/papers....
Fair AI isn’t a data problem you can clean your way out of. Unfairness is baked into the math itself, the task, the finite sample, the model. Even with perfect, unbiased data, a fair-and-accurate model can be provably impossible. Stop blaming the dataset. arxiv.org/pdf/2606.178...
𝑻𝒉𝒆 𝑫𝒊𝒈𝒊𝒕𝒂𝒍 𝑩𝒓𝒂𝒊𝒏: new version out. papers.ssrn.com/sol3/papers.... It reads an entire legal corpus as a knowledge graph you query in plain language. What that buys you depends on who you are:
I’ve rarely seen regulatory capture argued for (by Anthropic) this openly. Build an “FAA for AI” that can veto your competitors’ releases, staff it with government “approved” evaluators..... and justify the whole apparatus with existential-risk fan fiction. darioamodei.com/post/policy-...
I am HIRING three phd candidates. www.academictransfer.com/en/jobs/3617... A reshare would be most appreciated!!
3/ Our answer is contrarian. Computation is not the enemy of innovation competition. Done right, it gives innovation a more robust role than the old case-by-case approach ever did. Nothing is automatic. The challenge is ours to get wrong. Read it 👇 papers.ssrn.com/sol3/papers....
2/ Here is the worry. Every tool so far is built for what counts cleanly. Prices, quantities, document volume. Innovation does not count cleanly. There is no economic theory of it. So the machines get sharper while the blind spot grows. Or does it?
1/ From the US FTC to Saudi Arabia, antitrust agencies are building computational tools. Bid-rigging detectors. Contract scrapers. LLMs for spotting illegal information exchanges. Antitrust 3.0 is here. The question is what gets left behind. #computationalantitrust
Most automation has nothing to do with AI. A new global atlas of 124 countries shows low-income economies face the highest share of labor-substituting automation. Plain rule-based software does the bulk of it. China tops the world at 61.6% of tasks exposed. South Sudan sits at 3.3%.
IMPORTANT ONE! "Fewer innovations can mean higher welfare." The new piece by Callander, Lambert & Matouschek in @networklawreview.bsky.social explains when competition policy should stop counting products and start asking what kind they are. www.networklawreview.org/rugged-lands...
How are courts using AI? What limits should apply? Very pleased to be receiving Judge Zilgavis (CJUE) 👇 🗓 June 16, 12:00 to 13:00 Register: alti.amsterdam/event/peteri...
New paper. ChatGPT availability had essentially zero effect on US high school test scores in 2023-2024. Both the doomers and the boosters were wrong.