```
When an enterprise technology team sits down to evaluate a large language model for integration into their stack, the label "open source" on a model card carries significant weight. It implies transparency, reproducibility, and the freedom to modify and redistribute — values that underpin compliance frameworks and procurement decisions across the industry. But increasingly, that label is being applied to AI models that do not truly deserve it.
The openwashing problem
A growing number of AI models marketed as "open source" fall short of the Open Source Initiative's (OSI) longstanding Open Source Definition (OSD). These models may be freely downloadable, and their weights may be publicly available — a practice commonly described as "open weight" — but they lack key components such as full training data disclosure, complete source code for replication, or licensing terms that meet established open-source standards.
The distinction matters. An open-weight model allows users to download and run inference, but without access to training data, fine-tuning methodology, and reproducible build processes, the model cannot truly be studied, audited, or independently reproduced. For organisations that depend on software transparency for regulatory, security, or strategic reasons, this gap is not academic.
MOT: grading openness across multiple dimensions
As IBM's Arnaud Le Hors has discussed, assessing the real openness of an AI model is a non-trivial task. Models are not monolithic artefacts like traditional software packages. They comprise multiple components — training data, code, model architecture, weights, and licensing — each of which can vary independently along a spectrum of openness.
To address this complexity, the Model Openness Tool (MOT) has been developed as a structured evaluation framework. Rather than producing a simple binary verdict of open or closed, MOT examines individual components of a model and assigns granular assessments across dimensions such as data availability, code transparency, documentation, and licence permissiveness.
This multi-faceted approach reflects the reality that a model might publish its weights under a permissive licence while keeping its training data proprietary — technically "open" in one sense, but falling well short of the full transparency that the open-source community expects.
Where the debate stands
The development of MOT arrives amid a broader, ongoing debate within the OSI and the wider open-source community about what "open source" should mean in the context of AI. The traditional OSD was written with conventional software in mind, and its application to models trained on vast datasets raises questions that the original definition did not anticipate.
The OSI has been working toward a formal Open Source AI definition, a process that has generated significant discussion about where to draw the line. Should a model qualify as open source if its training data cannot be redistributed due to licensing constraints? What about models where the architecture is public but the specific training pipeline is not?
Tools like MOT offer a pragmatic response to these unresolved questions. By providing a standardised framework for evaluating model components individually, they give practitioners a way to make informed assessments today, even as the formal definitional debate continues.
Why it matters for practitioners
For technology professionals evaluating AI models — whether for deployment, integration, or research — the proliferation of "openwashing" creates real risk. Selecting a model based on an incomplete understanding of its openness can lead to unexpected licensing constraints, audit failures, or dependencies that cannot be independently verified.
The availability of an objective evaluation tool helps cut through marketing language and provides a factual basis for comparison. As the AI ecosystem continues to expand rapidly, with new models appearing weekly, the ability to systematically assess openness is becoming as important as evaluating performance benchmarks.
The conversation around open source in AI is far from settled, but the emergence of tools like MOT signals a maturing approach to a problem that the industry can no longer afford to ignore.
當企業技術團隊坐下來評估一個大型語言模型是否適合整合到其技術架構時，模型卡上的「open source」標籤具有重大意義。它意味著透明度、可複製性，以及修改和再分發的自由——這些價值支撐著業界的合規框架與採購決策。然而，這個標籤正越來越多地被用在那些名不副實的AI模型上。
開源洗白問題
越來越多被標榜為「open source」的AI模型，實際上未能符合開放原始碼促進會（OSI）長期以來的《開放原始碼定義》（OSD）。這些模型或許可以免費下載，其權重也可能公開取得——這種做法通常被稱為「open weight」——但它們缺乏關鍵組件，例如完整的訓練數據披露、用於複製的完整原始碼，或符合既定開源標準的授權條款。
這個區別至關重要。一個開放權重模型允許用戶下載並運行推理，但若無法取得訓練數據、微調方法和可複製的 build 流程，該模型就無法被真正研究、稽核或獨立複製。對於那些因法規、安全或戰略原因而依賴軟件透明度的組織而言，這個差距絕非學術問題。
MOT：多維度評估開放程度
正如IBM的Arnaud Le Hors所討論的，評估一個AI模型的真實開放程度並非易事。模型不像傳統軟件包那樣是單一的成品。它們包含多個組件——訓練數據、程式碼、模型架構、權重和授權條款——每個組件都可以在開放程度的光譜上獨立變化。
為應對這種複雜性，開發了「模型開放程度工具」（MOT）作為結構化的評估框架。MOT並非簡單地給出開放或封閉的二元判斷，而是檢查模型的各個組件，並在數據可用性、程式碼透明度、文件說明和授權條款寬鬆度等多個維度進行細緻評估。
這種多角度的方法反映了一個現實：一個模型可能在寬鬆的授權條款下發佈其權重，同時將其訓練數據保密——在某種意義上技術上是「開放的」，但遠未達到開源社群所期望的完全透明。
當前的爭論焦點
MOT的開發正值OSI及更廣泛的開源社群內部，就「open source」在AI背景下應意味著什麼展開一場更廣泛、持續的辯論。傳統的《開放原始碼定義》是針對常規軟件編寫的，將其應用於在海量數據集上訓練的模型，引發了原始定義未曾預料到的問題。
OSI一直在致力制定正式的「Open Source AI」定義，這一過程引發了關於界線應劃在何處的激烈討論。如果一個模型因授權限制而無法重新分發其訓練數據，它是否有資格被稱為開源？對於那些架構公開但具體訓練 pipeline 不公開的模型又該如何看待？
像MOT這樣的工具，為這些懸而未決的問題提供了一種務實的回應。透過提供一個標準化框架來分別評估模型組件，它們讓實踐者即使在正式的定義辯論仍在繼續之際，也能夠做出有依據的評估。
對實踐者的意義
對於評估AI模型的技術專業人士——無論是為了部署、整合還是研究——「開源洗白」現象的氾濫帶來了實實在在的風險。基於對模型開放程度不完整的理解來選擇模型，可能導致意外的授權限制、稽核失敗，或產生無法獨立驗證的依賴關係。
客觀評估工具的出現有助於穿透行銷語言，為比較提供事實依據。隨著AI生態系統持續快速擴張，每週都有新模型出現，系統性評估開放程度的能力正變得與評估性能基準一樣重要。
圍繞AI開源問題的討論遠未塵埃落定，但MOT等工具的出現，標誌著業界正以日趨成熟的方式處理一個不容再忽視的問題。