But Is It Open?The online edition
Type & theme

Make yourself comfortable

Page color
Text size
20 px
Reading width
Ask about this chapter ↗

Your settings and place stay in this browser.

Conclusion · Timothy O’Brien

Ask Before You Download

4 minute read14 source notesFree online edition

I grew up in the 80s and early 90s. In high school I skipped prom because I was so excited about Gopher. A year later I was using HTTP and HTML. You could tell something big was changing, even if you couldn’t say exactly what.

That memory needed one correction while I was writing this book. Gopher didn’t precede the web. It came out of the University of Minnesota in 1991, around the time the web was getting started at CERN. For a while, Gopher was ahead.1 Then, in February 1993, the university announced license fees for commercial Gopher servers.2 Bob Alberti, one of its developers, later said: “That socially killed Gopher.”3

Two months later, CERN put its web software into the public domain. Mosaic had been released that January, and through 1993 and 1994 the web gathered momentum.45 The licensing decision wasn’t the only reason Gopher lost ground. It helped change how people saw a system just as an alternative was spreading.

I remember the excitement more clearly than the terms. The terms helped decide what lasted.

Which layer stays open

We’re in another period when useful tools arrive faster than most people can understand the arrangements around them. Anthropic, OpenAI, and other companies have built remarkable systems. The people using them can still be dependent on policies and services they cannot change.

That dependence reaches beyond the model. What happens to a prompt after it leaves your machine? Which agreement governs retention, training use, and access? Can you change providers, run the system yourself, or inspect the part that failed? Different products from the same company can give different answers.

The comparison I return to is Netscape. Its browser dominated the early market and introduced proprietary extensions, but Netscape never owned HTTP or HTML. Those open protocols let other implementations compete. In 1998 Netscape itself released browser source code through Mozilla, which eventually led to Firefox.67 The open layer outlasted the browser fight.

I can’t tell you which layer of today’s AI systems will play that role. It could be model weights, training material, evaluation tools, interoperable data formats, or an interface that has not settled yet. The answer depends partly on what people choose to build and support now.

The evidence gives reasons to work on several layers. Stanford’s 2025 Foundation Model Transparency Index reported an average score of 40, down from 58 in 2024, with developers particularly opaque about training data and compute.8 OLMo demonstrates end-to-end release of model-development resources. Tülu demonstrates the value of publishing post-training recipes, while its cited Llama-based model retains the base model’s license and undisclosed pretraining corpus.9 Those are different contributions. Calling both simply “fully open” would lose the distinction this book has tried to preserve.

Peter Mattson of MLCommons put the importance of training material plainly: “Data is what determines the behavior of your system.”10 MLCommons’s safety working group has asked companies to share tests they otherwise run internally.11 Public artifacts make independent work possible; complete methods and usable data make that work more informative. The question is what another person can actually check.

Keeping the question useful

The Open Source Initiative matters because a label needs a boundary people can refer to. Its Open Source AI Definition emerged from a contested process, and disagreements about its requirements continue.12 Those disagreements deserve specific answers about permissions, code, parameters, and data information. A company should be able to show how its release meets a definition rather than relying on the word open in an announcement.

The same discipline applies to this book’s profiles. They are assessments of specified releases, with evidence and limits. They can help a reader find the missing permission or unpublished training resource.13 They cannot make the deployment decision, certify safety, or predict what a court will enforce.

There will be good proprietary products built on open foundations. There already are. Their usefulness doesn’t remove the need for foundations that remain available to competitors, researchers, and the people affected by a system’s decisions.

What participation looks like

An open future requires work beyond downloading. Read the terms before a deployment becomes a commitment. Report the defect with enough evidence that someone can act on it. Send a useful fix back. If your organization relies on a project or dataset, ask what keeps it maintained and help fund that work.

Participation also means looking at who can influence a release. A public repository may welcome corrections while keeping every consequential decision inside one company. A foundation may have formal independence while relying heavily on one employer’s engineers. Read the process as closely as the license.

In 1998, the new open source label described freedoms around software code.14 AI systems now make us ask about code, weights, training material, evaluations, and the services built around them. Each can have a different owner and a different set of permissions. A useful answer has to name the layer.

Before you download, ask what is available and what is missing. Ask what you may change and redistribute, what you can verify, and who can participate in the next release. Then ask what survives if the sponsor changes its mind.

Gopher was exciting enough to keep me home from prom. Excitement was never the guarantee. The permissions, the people, and the work they could carry forward were what mattered next.

Notes

  1. Christopher (Cal) Lee, “Where Have All the Gophers Gone? Why the Web Beat Gopher in the Battle for Protocol Mind Share,” University of Michigan School of Information, 1999 (revised 2008), https://ils.unc.edu/callee/gopherpaper.htm. Lee dates Gopher’s creation at the University of Minnesota to 1991 and CERN’s web to 1990, and reports Gopher traffic still ahead of the web through September 1993. Tim Gihring, “The rise and fall of the Gopher protocol,” MinnPost, August 2016, http://www.minnpost.com/business/2016/08/rise-and-fall-gopher-protocol, describes Gopher’s lead in users over the same period. ↩

  2. Wikipedia, “Gopher (protocol),” accessed 15 November 2026, https://en.wikipedia.org/wiki/Gopher_(protocol); secondary source, citing JQ Johnson’s 25 February 1993 post “gopher licensing” to the comp.infosystems.gopher newsgroup. MinnPost (see the next note) places the fee announcement at GopherCon ’93 and quotes the team’s defensive letter to users of March 1993. ↩

  3. Bob Alberti, quoted in Tim Gihring, “The rise and fall of the Gopher protocol,” MinnPost, August 2016, http://www.minnpost.com/business/2016/08/rise-and-fall-gopher-protocol. Alberti was a member of the original Gopher team. ↩

  4. Naomi Dinmore, “30 years of a free and open Web,” CERN, 30 April 2023, https://home.cern/news/news/computing/30-years-free-and-open-web. CERN released the World Wide Web software into the public domain on 30 April 1993. ↩

  5. NCSA, “A Legacy of Innovation,” https://www.ncsa.illinois.edu/about/history/ (Mosaic, January 1993); Christopher (Cal) Lee, “Where Have All the Gophers Gone?”, https://ils.unc.edu/callee/gopherpaper.htm (Gopher ahead in September 1993; sharply diverging growth around mid-1994). These sources do not establish a single end-of-1993 date when the web had replaced Gopher. ↩

  6. Wikipedia, “Browser wars,” accessed 15 November 2026, https://en.wikipedia.org/wiki/Browser_wars; secondary source. It reports that Netscape Navigator dominated the browser market by 1995 and describes the proprietary HTML tags Netscape and Microsoft added during the first browser war. ↩

  7. Netscape Communications, “Netscape Announces Plans to Make Next-Generation Communicator Source Code Available Free on the Net,” press release, 22 January 1998, reproduced in Mitchell Baker, “January 22, 1998: the Beginning of Mozilla,” 22 January 2008, http://blog.lizardwrangler.com/2008/01/22/january-22-1998-the-beginning-of-mozilla/. On the path from Netscape’s open-sourced code through the Mozilla Foundation to Firefox 1.0 (November 2004), see Wikipedia, “Browser wars,” https://en.wikipedia.org/wiki/Browser_wars. ↩

  8. Alexander Wan, Kevin Klyman, Sayash Kapoor, et al., “The 2025 Foundation Model Transparency Index,” Stanford CRFM, arXiv:2512.10169, 11 December 2025, https://arxiv.org/abs/2512.10169; index site at https://crfm.stanford.edu/fmti. ↩

  9. Ai2, “Olmo 3: Charting a path through the model flow to lead open-source AI,” November 20, 2025, https://allenai.org/blog/olmo3; Ai2, “Scaling the Tülu 3 post-training recipes to surpass the performance of DeepSeek V3,” January 30, 2025, https://allenai.org/blog/tulu-3-405b; Tülu 3 405B model card, https://huggingface.co/allenai/Llama-3.1-Tulu-3-405B (Llama 3.1 Community License; base pretraining corpus not disclosed). Open post-training resources do not make that base model an end-to-end open-source release. ↩

  10. MLCommons, “Peter Mattson Joins Leaders to Discuss AI Innovation at Asia Tech x Conference in Singapore,” June 6, 2025, https://mlcommons.org/2025/06/atx-panel. Mattson spoke at the panel the previous month. ↩

  11. MLCommons, “MLCommons Announces the Formation of AI Safety Working Group,” 26 October 2023, https://mlcommons.org/2023/10/mlcommons-announces-the-formation-of-ai-safety-working-group. The quoted words are from the announcement’s own text. ↩

  12. Open Source Initiative, “The Open Source Initiative Announces the Release of the Industry’s First Open Source AI Definition,” 28 October 2024, https://opensource.org/blog/the-open-source-initiative-announces-the-release-of-the-industrys-first-open-source-ai-definition; definition text at https://opensource.org/ai/open-source-ai-definition. The release quotes OSI Executive Director Stefano Maffulli calling the process “a difficult journey” marked by “differing opinions.” ↩

  13. But Is It Open? companion dataset and methodology, edition snapshot dated November 15, 2026, https://github.com/butisitopen/data/tree/main/snapshots/2026-11-15. Later release profiles and corrections are maintained at https://butisitopen.com/is-it-open/. ↩

  14. Open Source Initiative, “History of the Open Source Initiative,” https://opensource.org/about/history-of-the-open-source-initiative. The label was coined at a 3 February 1998 strategy session in Palo Alto, and OSI was founded later that month. ↩

From But Is It Open? by Timothy O’Brien (Discursive, 2026). © 2026 Timothy O’Brien. All rights reserved.

Conclusion4 min leftContents →

The book, chapter by chapter.

Source note

Go to the notes & sources ↓

The vocabulary.

A short guide to the book’s recurring terms. Follow each chapter for the argument and its sources.

Open weights
A release makes a trained model’s learned parameters available. The accompanying terms decide what you can do with them. Access to weights alone does not provide the training data, code or recipe. Read the chapter →
Copyleft
A way of using license conditions to preserve recipients’ freedoms when covered software is passed on. The obligations depend on the license and on what you distribute or make available. Read the chapter →
Permissive licenses
Licenses such as MIT and Apache 2.0 grant broad reuse rights with comparatively few conditions. Notice obligations still apply, and a software license does not automatically cover a release’s data or other components. Read the chapter →
Reproducibility
The ability to reconstruct and examine how an artifact was produced using the resources that are published. Running an existing checkpoint is a different achievement from rebuilding its training. Read the chapter →
Governance
Who can participate, who makes decisions, and how authority over a project or release is shared. A right to use the code does not by itself create a voice in upstream decisions. Read the chapter →
Training data
The examples used to train a model. Publishing the data, documenting its provenance and describing how it was filtered are distinct kinds of disclosure; privacy and other rights may limit what can be shared. Read the chapter →

A passage worth sharing.

Select a passage in the chapter, or share the one you’re reading.