HN 日本語サマリー

← 一覧へ戻る
科学・技術

最小のノペルト多面体:セプテペルト

Smallest possible nopert polyhedra: the Septopert (github.com)

44 pointsby RebelPotato6 コメント

要約

本記事は、7つの頂点を持つ凸多面体「セプテペルト」を紹介しています。これは「ノペルト」(ルパートではない多面体)として可能な最小のものです。セプテペルトは、その幾何学的特性、座標、対称性、そしてLean定理証明器を用いた数学的証明について詳述されており、特に「立方体の角」と「正確なトライフォース」というユニークな特徴が証明を複雑にしています。

全文翻訳

septopert.md

Tom 7, 2026年11月7日

セプテペルトは、7つの頂点を持つ凸多面体であり、ルパート(「ノペルト」)ではありません。7つの頂点は、ノペルトとして可能な最小数です。このページで説明されている標準的なセプテペルトは、ホロタイプであるSeptopert Classic、SeptopertM18、およびSeptopertM12を含む、類似のノペルト群の中で最も優れたバージョンです。その整数座標は、このファミリーのノペルトとして可能な限り最小です(「最小高さ」を参照)。このページでは、正確な定義、Leanがどのようにチェックしたか、そして結果を再現する方法を示します。

Leanのエントリーポイント:proof/Septopert.lean;主要な定理はproof/Septopert/Integer.leanにあります。

立体

[画像: セプテペルトが自信を持って回転している]

ここに正確なSTLファイルがあります。GitHubでは、これをクリックするだけでブラウザ上で回転させることができます。

セプテペルトは、これらの点(その頂点)の凸包です:

(-7,  7,  7)
( 7, -7,  7)
( 7,  7, -7)
( 3, -4, -4)
(-4,  3, -4)
(-4, -4,  3)
( 7,  7,  7)

これは非常に「7」にちなんだ立体です:7つの頂点、座標は±7と3、4(3 + 4 = 7)、頂点エッジの長さは14、基底エッジは14√2、内側の三角形エッジは7√2です。Tom 7によって発見されたにもかかわらず、この「7」という数字は単なる数学的な偶然です!ジャックポット!

Leanの記述は、まさにこれらの整数点(最大の|v| = √147 ≈ 12.12)に関するものです。この立体の任意のスケール変換または剛体変換もノペルトです。(内部的には、証明は実際には古い機械の一部を再利用し、異なるフレームとスケールで動作し、その後補題を使用して上記の頂点に対するステートメントを証明します。)

  • 他のメンバーと同様にC3v対称性:(1, 1, 1)に沿った3倍軸と3つの鏡面。頂点0〜2は(−7, 7, 7)の置換、頂点3〜5は(3, −4, −4)の置換、頂点6は7·(1, 1, 1)です。
  • 10個の三角形の面、ファミリーの他のメンバーと同じ10個(頂点リスト015、016、024、026、045、123、126、135、234、345);頂点次数は5,5,5(頂点0〜2)、4,4,4(3〜5)、3(頂点6)。組み合わせ的には、ジャイロ延長三角錐です。
  • 2つの特別な特徴(両方とも正確)があり、これが証明をより難しくしています:
    • 立方体の角:頂点6は、SeptopertM18と同様に立方体の角です。エッジ6–0、6–1、6–2は互いに直交しているため、3つの頂点二面角は正確に90°です。
    • 正確なトライフォース:3倍軸に沿った正射影(原点を通る、(1, 1, 1)に垂直な平面上)では、頂点3の画像は頂点0の画像のちょうど−1/2倍です(以下同様)。したがって、内側の三角形3–4–5は、基底0–1–2の medial triangle に投影されます。あるいは、頂点4は0–2の中点に投影されるため、面0–2–4の平面は軸に平行であり、面0–2–4は面3–4–5(法線が軸)に垂直です。これは2番目の直角であり、最初の次数が消えるビューの2番目のファミリーを提供します。
エッジ情報

正確な算術(整数面の法線;上記の整数座標での長さ)。

[画像: セプテペルトの2つのビューの線画、色付きのエッジ付き]

(注:これはセプテペルトファミリーのどのメンバーにも使用される模式図であり、正確ではありません。)

色説明エッジ二面角エッジ長
[画像: 青]ピラミッドの頂点6–0, 6–1, 6–290° (正確)14
[画像: 黄]ピラミッドの基底0–1, 0–2, 1–2144.735610°14√2 ≈ 19.798990
[画像: 緑]トライフォース外側0–4, 0–5, 1–3, 1–5, 2–3, 2–4108.316731°√146 ≈ 12.083046
[画像: 赤]トライフォース中心3–4, 3–5, 4–5141.057559°7√2 ≈ 9.899495

定義

proof/Septopert/Integer.leanにて:

theorem isNopert_intVerts (h : CertificatesValid) : IsNopert intVerts

intVertsは上記の7つの整数点の集合です(septopertはそれの別名です)。

IsNopert、IsRupert、および信頼モデルはSeptopertM18と同じです。CertificatesValidは、証明データ(4つの解テーブル、ミラーキャップ証明、および3つの特殊キャップ:アークキャップ、軸キャップ、Pキャップ)がLeanの妥当性チェックを通過したことを示します。同じIsNopertが最小性の結果によって使用されており、セプテペルトはproof/ProofOfMainTheorem.leanにおける7頂点の証拠です。

証明書は、立体の合同コピー(頂点、座標は64分の1:(−27,15,15)、(27,6,6)、(−13,−13,−13)およびその置換)で構築されます。短いブリッジ補題はそれを正確な相似(スケール3/64、軸を通る鏡映とそれに関する180°回転の合成、および平行移動)で整数座標にマッピングし、相似はルパート性を保存します。

証明されたことと検証されたこと

SeptopertM18と同様に、isNopert_intVertsはLeanのカーネルによってチェックされ、標準公理(propext、Classical.choice、Quot.sound)のみに依存します。データ(約732,000行のテーブル行とキャップ証明書)は、プログラムconstructSeptopertによってチェックされます。これはコンパイル済みコードと同じLean妥当性チェックを実行し、その後IsNopert intVertsを構築して以下のように出力します:

instantiated: IsNopert Septopert.intVerts (CapOk, ArcOk, AxisOk, PCapOk checked natively; tube and chart tables checked)

その他のセプテペルト

このページでは、整数座標が最小の標準的なセプテペルトについて説明しています。これは、正確な立方体の角と正確なトライフォースを持つ、ファミリーの端(我々の考えでは)に位置しています。

証明の過程で、ファミリーの他のいくつかのメンバーがノペルトであることを証明しました:

  • Septopert M12は同じ二重特殊な位置にあります。
  • Septopert M18は立方体の角を持っていますが、トライフォースは約3.7%ずれています。
  • Septopert Classicは一般的な位置にありますが、近い値です(二面角90.333818°、トライフォースから3.5%オフ)。

証明の仕組み

技術はHow the Nopert proofs workの通りで、SeptopertM18のアークキャップを使用しています。ビューカバーは13個のビューコードと16個のビューコード三角形を持っています(SeptopertM18:15個と22個)。チューブと5次元検索は、4つの小さなビュー領域を除くすべてを閉じますが、これらの各領域は、一次の余裕が消滅するため、ブローアップ証明書が必要です:

  1. ミラーキャップ:ミラー法線でのキャップ。ファミリーのすべてのメンバーと同様です。
  2. 立方体の角のアーク:SeptopertM18と同様。頂点エッジをまっすぐ下に見るビューUで終わるビューのアーク。そのエッジを通る面がエッジオンになり、頂点の直角によりすべての一次の余裕がゼロになります。これはアークキャップ(アークに沿ったウェッジチャートとUでのポイントキャップ)によって証明されます。
  3. 軸ビュー:3倍軸に沿ったビュー。ここではトライフォースが内側三角形の影を基底の medial triangle に正確に配置し、一次が再び消滅し、孤立したビューになります。これは軸周りのポイントキャップ(Cayley座標で1/46以内の回転;傾きμ·(s₁, s₂)、スピンμ²·s₃)によって証明されます。
  4. 面0–2–4のストリップ:すべての面はビューのグレートサークルに沿ってエッジオンになります。面3–4–5に垂直な面0–2–4の場合、直角により、ビューP(頂点の直角が立方体の角のアークに沿ってそうするように)から始まる円のストレッチに沿って一次が消滅します。ここでは、一次のフレキシビリティの2次元コーンがあります。これはPキャップによって証明されます:ストリップに沿ったウェッジ極座標チャートと組み合わせた終端ビューPでのポイントキャップ。

各ケースで、決定は二次に移行し、クリアランスは厳密に負であるため、証明が存在します(浮動小数点および記号的研究、証明の一部ではありません)。

最小高さ

ファミリーのメンバーを5つの座標値x = (x0, ..., x4)で記述します。点(x0, x1, x1)、(x2, x3, x3)およびそれらの置換、そして(x4, x4, x4)です。その高さは、これらの値の範囲の半分を切り上げたものです。セプテペルトの値は−7、7、3、−4、7(範囲14)なので、高さは7です。これより低い高さの整数メンバーはノペルトではありません。proof/SeptopertFamily.leanにて:

theorem not_isNopert_of_familyHeight_lt_seven (hvalid : LeastHeightValid) : 
    ∀ x : Fin 5 → ℤ, familyHeight x < 7 → ¬ IsNopert (familyVerts x)

このステートメントは、任意の符号を持つすべての整数タプルをカバーします。タプルを変換、スケーリング、および反転すると、ノペルトであることは維持されます(Leanで証明済み)。そのため、値が[0, 12]の範囲にある55,651個の標準的なタプルに還元され、それぞれにcheckSeptopertFamilyによってチェックされる小さな証明書があります(約4秒):

  • 161個は共面です(7つのすべての点を通る平面)。
  • 25,932個は、他の点の凸包内に点があります(明示的な凸重み)。したがって、最大6つの点が凸包をスパンし、最小性定理が適用されます。
  • 29,558個は正確なルパート通過を持ちます:整数四元数q(R(q)/|q|²として)からの回転、有理オフセット、および各内側頂点に対して、その影を厳密に含む外側三角形があります。

テーブルはproof/SeptopertFamily/data/leastheight.txt(2.7 MB)です。他の結果と同様に、これはカーネル定理(公理propext、Classical.choice、Quot.sound)に加えて、データのネイティブチェックです。

ステータス

2026年11月7日に検証済み:constructSeptopert(その後constructM7)は、ミラーキャップ(52チャート)、アークキャップ(146チャート、22,282リーフ)、軸キャップ(52チャート、3,384リーフ)、Pキャップ(78チャート、43,869リーフ)、チューブテーブル、および4つのチャートテーブル(406,726、124,781、108,449、92,514行)をチェックし、instantiated行を上記のように出力しました。ビジーなマシンの12コアで約23分かかりました(ステージ時間:ミラーキャップ196秒、アークキャップ723秒、軸キャップ16秒、Pキャップ300秒、テーブル残りの時間)。

その後、2026年11月8日に、このリポジトリ(コミット3d8fce61)の新規クローンから、アイドル状態の64コアマシンの60スレッドを使用して、ゼロから再現されました。3つの特殊キャップ証明書を含むすべてのステージが再生成されました。チャートテーブルは上記と同じ行数を持ち、再生成されたモデルファイルはコミットされたものと一致し、constructSeptopertは同じinstantiated行を出力しました:

./get-cc-lib.sh
(cd proof && lake exe cache get)    # Mathlib build cache
search/septopert/rebuild.sh --threads 60 --work work-fresh

検索はマルチスレッドであり、決定論的ではないため、他の実行では詳細が異なる場合があります。ステージ:model、tube、idcap、tubepack、cap、arccert、axiscert、pcapcert、search、pack、verify(--from STAGE、--only STAGE、--work DIR;デフォルトはmodelから開始)。測定されたステージ時間(合計31分):

ステージ時間
model9 秒
tube35 秒
idcap781 秒
cap24 秒
arccert96 秒
axiscert2 秒
pcapcert93 秒
search69 秒
pack41 秒
verify727 秒

作業ディレクトリは274 MBでした。

データ

証明書データは機械出力であり、CC0を意図しています:

  • proof/Septopert/data/axis-cap.cert(148 KB)およびproof/Septopert/data/pcap-wp.cert、pcap-u.cert(それぞれ349 KBと1.5 MB):軸キャップとPキャップはリポジトリ内にあります。
  • arc.cert:916 KB、アークキャップ証明書(146チャート、22,282リーフ)。
  • cap-1_16.cert:408 KB、ミラーキャップ証明書。
  • tube_packs/:944 KB、16個のIDチューブテーブルとmanifest.txt。
  • packs/:37 MB、64個のソリューションパック、4つのチャートと16個のビューコード三角形のそれぞれ1つ(約732,000行)。

リポジトリにない4つは、約6.8 MBに圧縮されます(tar czf)。


[2] タイトル: Generative AI Usage in the Games Industry
URL: https://www.gameworkers.co.uk/ai-manifesto/
本文:

On Generative AI Usage in the Games Industry

The following manifesto was written by members of the IWGB Game Workers' Branch in collaboration with members of Game Workers Unite Ireland of the Republic of Ireland, CSVI of Spain, and the FIOM-CGIL Milan Work Council

Introduction

Our Aims

We, a collective of video game unions and work councils from Europe, write this manifesto to raise awareness of the harms of generative AI (genAI) use within the video game development sector.

GenAI is currently harming the games industry, its workers and the games they produce. Left unchecked, this harm will continue to grow and impact more workers, more games, and more players, even more acutely.

Our goal is to bring to light the risks and dangers genAI poses to the workforce, as well as to set guidelines and standards for games industry workers, stakeholders, and regulators, suggesting actions which can be taken to reduce – and ideally eliminate – these dangers.

Scope

This manifesto discusses genAI in the context of the video game development sector. These uses include, but are not limited to:

  • The creation of assets included in a released game which are experienced by players – for example: art, sound, code, and dialogue assets.
  • The creation of materials which are used as part of the process of development but which players won’t experience when playing the released game – for example: early concept art, reference materials, pre-translations, etc.

By “genAI” we mean a specific type of AI used to generate text, pictures, videos, or music in response to limited inputs (usually just text “prompts”) using LLMs and other energy-intensive models. The wider harms of genAI won't be covered here, but they constitute additional extremely compelling reasons to resist the use of genAI. These include:

  • Severe harm to the environment (Zewe, 2025)
  • The destruction of communities (Nguyen, 2025)
  • The exploitation of the Global South (F.Regilme, 2025)

This document won’t go into detail about how different genAI models work, nor discuss the comparative quality of their output.

The focus of this document is on games industry workers (us!), how we are and will be impacted, and how we can resist.

The Impact of GenAI

GenAI Creates Worse Working Conditions

  • GenAI is both directly and indirectly replacing game industry workers.
  • Many game workers are being forced to use genAI, either explicitly or implicitly.
  • Working with genAI is less desirable, fulfilling, and educational than the same work without genAI.
  • Over time, working with genAI leads to worker fatigue, cognitive decline and burnout.
  • Using genAI removes a sense of ownership from workers.
  • GenAI is causing a marked decline in junior positions within the industry, leading to a lack of new talent.

Proponents of generative AI claim that it is a tool of empowerment for workers. For that to be true, every worker would be free to choose if, when, and how to use it. We would also need to trust that every employer would use the promised productivity improvements to benefit their workers.

In practice, this is not what we are seeing.

In many workplaces today, workers are being forced to use genAI ‘tools’, sometimes in very specific ways (Plunkett, 2025). Some companies do not explicitly impose the usage of genAI, but set expectations which cannot be reached with or without its use. Allowing the choice of usage, while also creating expectations that the worker should be able to keep pace with automated procedures, is no choice at all (Shadmi & Merchant, 2025).

When workers do use genAI, they find work more onerous than before. Many of the tasks related to genAI work are more tedious and less fulfilling, since the work becomes less about exercising the skills workers have built and invested in, and instead pivots to a plethora of small, parallel tasks which amount to supervising genAI and bug-fixing its output.

Workers using genAI have reported ‘AI brain fry’ and reduced focus (Pearl et al., 2025). Although gains to productivity are possible, these tend to be short term. Employees often initially work faster, take on larger tasks, work overtime, and ignore breaks, due to working with genAI feeling 'more like chatting' (Rangantathan & Ye, 2026). However, once this novelty declines, so does productivity, with workers reporting cognitive fatigue, burnout, and inhibited decision making.

A recent study (Shaw & Nave, 2026) showed that participants who used genAI frequently adopted results that were systematically incorrect, especially under time pressure but also in the face of incentives and feedback. The researchers were able to prove that users of genAI habitually delegate their agency and offload decision making to an external party with relative ease. They call this process ‘cognitive surrender’ (Shaw & Nave, 2026).

The use of genAI also removes a sense of ownership from workers, who become alienated from the results of their labour: the things they produce no longer feel like they have the workers’ stamp on them, and therefore they can no longer take pride in them.

GenAI is disproportionately used for work which would traditionally be given to junior workers, and is hence directly reducing the amount of junior positions available within the industry. This is worsening an already existing problem for newcomers to the industry, who are finding it harder than ever to find work, but it is also a loss of potential fresh perspectives, and is staggeringly short-sighted when it comes to the industry’s sustainable future.

GenAI has a profound negative impact on workers by reducing the availability and appeal of work. Games are best made by people, and people do their best when they are healthy and connected to their work, refining their skills and accumulating real experience.

GenAI Harms Relations Between Workers and Employers

  • GenAI is sold to employers as an alternative to workers.
  • GenAI often increases pressure on workers when used by employers to set policy, timescales, etc.
  • GenAI can be used by employers to abdicate responsibility.

GenAI’s promise to employers is to more fully automate processes, to enable them to be completed with fewer workers. This directly reduces the amount of jobs available for game workers. In some cases, workers have even been forced to train the models which then replaced them, leaving them unemployed.

GenAI interfaces are engineered to be human-like and eager-to-please, to prioritise keeping the user engaged, even if engagement comes at the cost of providing them with incorrect information. GenAI’s predisposition towards this sycophancy and lack of real-world context can lead to an increase in unreasonable, damaging requests when used by employers to set policy, timescales, workloads, etc., as well as to task expansion and a general degradation of conditions (Cheng et al., 2026).

In addition, when employers engage with human-like genAI models which never tell them ‘no’, they become less receptive to criticism and pushback from other workers.

GenAI can also obfuscate and confuse employers’ accountability, in two main ways. Firstly, employers may shift their own accountability onto these platforms, blaming an error on the “tool”, rather than accepting responsibility themselves. Secondly, they may blame workers for issues which arise through the use of genAI – which the employers have mandated.

Any time genAI is used without exacting human verification, the possibility of error increases. Exacting human verification takes time, which negates the promise of genAI automation, and therefore many employers are choosing to skip this step, either directly or indirectly by imposing time constraints on workers which make detailed verification impossible. Workers are then blamed for failures in the genAI’s output.

Employers see genAI as a way of cutting labour costs, employing fewer workers than they would need to without it. GenAI companies ingratiate themselves with employers, by promising the impossible. Moreover, mistakes made by genAI are often blamed on the workers, rather than the employers mandating its use.

GenAI Appropriates the Work of Workers

  • GenAI models are often trained on stolen third-party data, with employers failing to credit or compensate the original creators.
  • Where genAI models are trained on internal data, employees/creators are likewise rarely credited or compensated.

GenAI models are mostly trained on stolen datasets, and, as stated by the owner of a genAI company, ‘impossible to create’ without stealing data (Milmo, 2024). Employers often feel entitled to automatically include their employees’ contribution into their internal dataset without consent and/or extra compensation.

This point does not require much explanation, but please do not let the relative smallness of this section undermine its importance: stealing the creations of others is wrong, creating very similar versions of those creations is wrong, and using those similar versions to generate wealth which the original creator does not share in is wrong.

GenAI does not create, it appropriates. Content output by genAI has its origins in the work of humans, who are rarely compensated for that output.

GenAI Perpetuates Discrimination

  • Most genAI models are trained on datasets that carry an intrinsic and exclusionary Western bias that limits diversity and marginalises non-dominant languages, cultures, and ethnicities.
  • GenAI content averages out data used to train the AI, leading to a homogenisation of culture.

GenAI models are trained on vast swathes of historical data which is often unvetted and without context, as they aggregate large amounts of data and look for commonalities to generate a response without any direct connection to reality. The use of old data can reflect outdated opinions and values which we would be horrified at today, as well as generate middle of the road solutions.

The predominance and absence of certain types of people in historical data can also reinforce their predominance or absence in the output of genAI models today: for example, the predominance of white male doctors in historical medical literature may lead genAI models to assume that a referenced doctor is white, or to create a white male character when asked to create a doctor.

Within the games industry, this has two main types of effect: it can increase discrimination within workplaces, for example, when used in hiring practices, and it can reduce the range of representation within created media, leading to more homogenous characters and viewpoints in the games we create and play.

GenAI reproduces both historical and contemporary biases and can lead to both implicit and explicit discrimination, including but not limited to racism, sexism, homophobia, transphobia, etc.

GenAI Has Hidden Costs

  • GenAI’s output is unreliable and often incorrect.
  • GenAI providers obfuscate the cost of genAI, its inner workings, and its limitations.
  • GenAI providers aim to silo and monopolise knowledge to sell it back to corporations.

GenAI providers obscure a significant portion of their associated costs (to the environment, mental health, etc.), which impacts the ability of workers to make an informed choice on whether and how to use them (Nguyen, 2025).

GenAI platforms also disincentivise education on their limitations. Average usage of genAI goes down with increased understanding of how genAI actually works behind the scenes (Tully et al., 2025, 20).

GenAI providers sell employers a seductive image of the future, where production can be fully automated. However, in actuality, genAI platforms are gatekeeping transferable skills (Shadmi & Merchant, 2025). Once a model is trained, the ‘owner’ (usually a large conglomerate) of the trained AI has a uniquely powerful position and complete ownership over skills that were previously owned and shared by workers.

This is bad for workers, but it is also bad for employers who outsource their knowledge and skills to third-party genAI providers, as once these companies have a monopoly on this previously institutional knowledge, their prices are unlikely to remain competitive.

GenAI providers aim to replace employers’ dependence on workers with dependence on their own models. Companies which buy into this promise while genAI usage is cheap may come to regret this when it is the only option and prices increase. Workers will regret this much sooner.

Recommendations

Given the present and future impact of genAI on the games industry, we make the following recommendations:

Recommendations for Workers

We understand that the industry is very unstable at present, and that can make it more difficult to speak out. Please do as much as you are comfortable with, and remember that we are stronger together. 

  • Avoid usage of genAI services whenever possible.
  • Make your voices heard.
    • Distance yourself from genAI usage.
    • When genAI content is presented to you, question its use and express dissatisfaction.
    • Talk to your colleagues about the risks and harms of genAI.
    • Make it clear to your employer why you do not want to use genAI.
    • Support other colleagues when they speak out about genAI.
  • Do not burn yourself out because of artificially inflated expectations. Do not be a hero by trying to work at an unrealistic pace.
  • Educate yourself on how to spot genAI content so that you can avoid accidentally using it in your own work.
  • Seek out information on your company’s genAI services from impartial sources, and share it with your colleagues.
  • Ask your local worker representatives to negotiate an agreement on genAI usage for your company. Discuss its contents with your colleagues.
    • This agreement could reinforce protections on refusing to use genAI.
    • You could ask to form a committee to evaluate/investigate violations on genAI usage.
  • Organise. Unionise. Do it now, not when it is too late and you are already under threat.

Recommendations for Employers:

  • Listen to your workers.
    • Involve the workforce in development of policy regarding genAI.
    • Workers should be the ones who decide when, where, and how to use it.
  • Listen to your customers.
  • Avoid the use of generative AI entirely.
    • Ensure that no worker is required to use genAI.
    • Disallow the use of genAI in the company code of conduct, even for early stages of development and prototyping.
    • If it is absolutely impossible to avoid usage, label/declare usage precisely.
  • The chain of accountability at work should be respected and managers should be held accountable for the output of their team.
  • Inform workers about the negative outcomes of using genAI. Information should come from impartial sources, not companies trying to sell their genAI services.
  • Schedule training for workers on how genAI actually works behind the scenes, its real limits and capabilities.
  • Do not use your employees’ work as training data. If you do:
    • Only use data after explicit consent.
    • Ensure fair compensation for reuse of their labor.
    • Allow them to withdraw their work from such datasets.

Recommendations for Players:

  • Support games made without genAI.
  • Be vocal about disliking genAI in games – leave comments, make videos, ask for change.
  • Educate yourself on how to spot genAI content so that you can avoid it.

Recommendations for Regulators:

The UTAW Tech Workers’ Inquiry have put together an incredible resource with suggestions for regulators which we wholeheartedly agree with:

https://techworkersinquiry.org/ai/

In addition to their points, we would also like to suggest that:

  • It should be forbidden to submit a person’s work to any genAI dataset without explicit permission and consent, and most importantly without fair compensation for reuse of labour.
  • Any acceptable submission must be for datasets that are transparent in how data is used, stored, and retained. Workers must be able to withdraw their content from such datasets at any time

Conclusion – GenAI Against Humanity

GenAI is worsening our working conditions. Our belief is that workers should be the ones who decide when, where and how to use it, as they are the ones who know how to do their jobs. No one likes being forced to do something that harms them, and a lot of workers are being forced to use genAI against their will.

We do not have to accept the use of genAI to make video games;

it is not inevitable.

If genAI was just a useful tool, as advertised, it would make our lives easier and we would have more free time, but the opposite is happening. The owners of these tools are aware of this and that is why they push it: to normalise it, to spread its use, and to ensure it becomes indispensable before raising prices. They are the ones in power, not workers, not even CEOs in video game studios.

We firmly believe that video games are nothing without the humans who make them – they are a way  for us to express ourselves, to share knowledge or a point of view with the world, to entertain all kinds of gamers. There is no intent in anything made by a genAI probabilistic engine. If no one put their skill, effort, soul, and time into making a game, why would anyone waste their time playing it?

In the current context, everything has to be data-driven: percentages, averages, time spent, productivity… But quality is subjective and cannot be measured, so companies disregard it. People want high-quality games, they want to spend their time playing a game whose developers poured everything into making it, because they are more enjoyable, as it happens with any other art form.

Years after the public release of genAI chatbots, their use has not been as widespread as their owners would like, but they have been deployed as an excuse for mass layoffs with the goal of pleasing shareholders. We should stop thinking about what is good for business, but instead think about what is good for society in general, for workers, for gamers. What good is it if a CEO earns millions every month if the working conditions at their studio are worsening, along with the video games they create?

No genAI will be able to replace the humanity in any work of art.

We are sick of shareholder value being the be-all and end-all of this industry. 

We are sick of watching five giant conglomerates just shuffling money around in a giant scam to be the winner of this never-ending hype contest.

We are sick of seeing people fired or forced to quit just to maximise profits. 

We are sick of being told to just accept that this is what is good for business.

GenAI is bad for workers and games. Workers, protect your interests. No games get made without our collective strength. Organise. Fight for your rights.

FAQ

  • What happens if most of my colleagues want to use genAI in some capacity? In other words, if a developer really wants to use genAI as part of their process, can they use it?

  • The short answer is yes. Obviously the other problems are still there, therefore you should decide as a collective if you want to use it, on the whole, given concerns around the environment, cost, ethics, law, honesty, the sanctity of art, etc. Our aim in this document is to educate people about the harms of genAI, so that both employees and employers are less inclined to use it.

  • What if the quality of the output is good, or good enough?

  • This doesn’t change the main focus of our manifesto. We want workers’ conditions to improve, and to organise against abuses and for better ownership.

  • Why are you talking about AI and not layoffs, RTO mandates, workplace toxicity, etc.?

  • While the issues caused by genAI are coming from the same place as these other adjacent crises, the genAI situation is developing rapidly and has practical knock-ons to every aspect of our industry. We’re still working on the other issues plaguing our industry, with old and new methods, and we encourage everyone to join us in that fight.

  • How do I know if I should get compensation for an AI-generated output? Is there even a way to find out if a specific data entry was important for a generated output? Or to remove a data entry from a set?

  • There are examples of research going in this direction like “WAGLE: Strategic Weight Attribution for Effective and Modular Unlearning in Large Language Models” (J Jia, J Liu, Y Zhang, P Ram, N Baracaldo, S Liu, 2024) and  “Losing is for Cherishing: Data Valuation Based on Machine Unlearning and Shapley Value” (Ma, Le ; Yang, Shirao ; Wang, Zihao ; Wang, Yinggui ; Wang, Lei ; Wei, Tao ; Zhang, Kejun, 2025), but the specific implementation is not important.

  • Why are you against the future?

  • We are not. We simply believe that workers have the right to shape their future. Being exploited and overworked is not a future we are willing to accept. Time will prove genAI ‘tools’ and their usefulness, let us not be trampled in the meantime. The future is not yet written.

References

Azar, J., Gine, M., & Sanz-Espín, J. (2025, December). The Wage Effects of Generative AI. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5842084

Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026, 3 26). Sycophantic AI decreases prosocial intentions and promotes dependence. Science. Retrieved July 31, 2026, from https://www.science.org/doi/10.1126/science.aec8352

Francis, B. (2025, 9 16). Devs are more worried than ever that generative AI will lower the quality of games. Gamedeveloper. https://www.gamedeveloper.com/business/devs-are-more-worried-than-ever-that-generative-ai-will-lower-the-quality-of-games

F.Regilme, S. S. (2025). Artificial Intelligence Colonialism: Environmental Damage, Labor Exploitation, and Human Rights Crises in the Global South. SAIS Review of International Affairs, 44(2), 75-92. https://muse.jhu.edu/article/950958

J Jia, J Liu, Y Zhang, P Ram, N Baracaldo, S Liu. (2024). Wagle: Strategic weight attribution for effective and modular unlearning in large language models. NeurIPS. https://arxiv.org/abs/2410.17509

Ma, Le ; Yang, Shirao ; Wang, Zihao ; Wang, Yinggui ; Wang, Lei ; Wei, Tao ; Zhang, Kejun. (2025, 5 1). Losing is for Cherishing: Data Valuation Based on Machine Unlearning and Shapley Value. eprint arXiv:2505.16147. https://ui.adsabs.harvard.edu/abs/2025arXiv250516147M/abstract

Milmo, D. (2024, 1 8). 'Impossible' to create AI tools like ChatGPT without copyrighted material, OpenAI says. The Guardian. Retrieved July 31, 2026, from https://www.theguardian.com/technology/2024/jan/08/ai-tools-chatgpt-copyrighted-material-openai

Nguyen, T. (2025, 07 1). What happens when data centers come to town? Ford School University Of Michigan. Retrieved July 31, 2026, from https://stpp.fordschool.umich.edu/sites/stpp/files/2025-07/stpp-data-centers-2025.pdf

Pearl, M., Lapa, R., & Merchant, B. (2025, 9 16). Artists are losing work, wages, and hope as bosses and clients embrace AI. Blood In The Machine. Retrieved July 31, 2026, from https://www.bloodinthemachine.com/p/artists-are-losing-work-wages-and

Plunkett, L. (2025, April 7). An Overwhelmingly Negative And Demoralizing Force’: What It’s Like Working For A Company That’s Forcing AI On Its Developers. Aftermath. Retrieved July 31, 2026, from https://aftermath.site/ai-video-game-development-art-vibe-coding-midjourney/

Ranganathan, A., & Ye, X. M. (2026, 2 9). AI Doesn’t Reduce Work—It Intensifies It. HBR. https://hbsp.harvard.edu/product/H0929G-PDF-ENG

Shadmi, K., & Merchant, B. (2025, 12 12). *I was forced to use AI until the day I was laid off.