はじめに
ChatGPT Plusユーザーとして、GPT-4.5使用時に感じていた「不透明な使用回数制限」「モデルの無通知切替」「実質的な性能限界」について、OpenAIに英語でフィードバックを送ったところ、型通りではない、明確で本質的な回答を受け取りました。
この記事では、その要望とOpenAIからの返答を全文公開し、今後の使用や改善提案の参考になる情報を共有します。
提出した要望内容(英訳)
✅ 1. Display remaining usage count for GPT-4.5
Currently, GPT-4.5 usage is limited to ~50 messages per week on the Plus plan.
- However, there is no indication in the UI about how many uses are left.
- Even the model dropdown does not show remaining quota.
Request: Please make the remaining usage count visible in the interface or at least notify the user when nearing the limit.
✅ 2. Notify users clearly when the model is automatically downgraded
- Upon hitting the GPT-4.5 usage cap, the model silently switches to GPT-4o mini (or another fallback).
- There is no popup, banner, or log entry indicating this change.
Concern: This is a silent downgrade of service quality. In consumer protection terms, this could be interpreted as a non-transparent product substitution, which may raise regulatory issues in certain countries.
Request: Notify users explicitly when an automatic model switch occurs due to quota limits or system load. The user has a right to know what they're interacting with.
✅ 3. Provide a session usage report (automatic summary)
At the end of each session, or daily, the following should be returned:
🧾 Model Usage Report (2025/05/11)
- GPT-4.5: 35 requests
- GPT-4o mini: 6 requests
- Automatic switches: 2 (last switch: 13:42)
- Remaining this week: 15 / 50
This would enable professional users to track usage accurately and ensure billing transparency.
✅ 4. Clarify and strengthen token limits in real use
While GPT-4.5 (turbo) is advertised as supporting up to 128,000 tokens, in practice:
- The model often breaks or switches around 10,000 characters (~2,500 tokens)
- This undermines trust and severely limits use in technical workflows involving long code files or logs.
Request: Either improve the true capacity, or clearly disclose the practical token limits per request/session.
⚠️ Emerging abuse risk: AI jamming via hidden long text
The current architecture allows malicious actors to embed hidden junk text in Markdown or PDFs (e.g., 10,000+ characters), causing:
- Response delays
- Model downgrades
- Context loss
This constitutes a form of AI jamming or denial of service via prompt overload. The system should include safeguards to detect or prevent such overload attempts.
🔚 Summary of Requests
- Show remaining GPT-4.5 usage clearly in UI
- Notify users explicitly on model switches
- Add automatic session reports summarizing usage
- Accurately document or improve GPT-4.5's practical token capacity
- Consider abuse risks from hidden long text and improve resilience
日本語で概要を示すと
✅ 1. 使用回数の可視化を求める(再掲)
現在、使用回数がUI上で一切表示されない。
ドロップダウンに「残り〇回」の表記が必要。
できれば「このセッションではGPT-4.5を〇回使いました」と終了時に返すべき。
✅ 2. モデルの自動切替に関する不透明性
使用制限到達時、GPT-4.5 → GPT-4o miniなどへの強制切替が通知なく行われる。
これは別モデル・別性能であり、契約の事実上の変更と見なされかねない。
🔥 消費者庁指摘の対象になる恐れあり:
「高性能モデルを使用しているつもりで、実は劣化版に自動で切り替えられていた」
これは明確な表示義務違反に近い。
✅ 3. セッションの自動レポート導入
セッション終了時に、以下のような情報を明示せよ:
🧾 利用モデルレポート(2025/05/11)
- GPT-4.5:35回
- GPT-4o mini:6回
- 切替:2回(最終切替:13:42)
- 今週の残り回数:15 / 50
✅ 4. 入力トークン限界の実質値が低すぎる問題
現在「128,000トークン対応」とうたいながら、実際には1万文字(=約2,500トークン)を超えると動作不安定・モデル切替・応答途切れが頻発する。
- GPT-4.5が「1万字で沈黙」するのでは高信頼な作業に使えない。
- GPT-4.5 turboを名乗るなら、その程度では沈黙しない設計にすべき。
⚠️ 技術的悪用のリスク
この仕様は逆に、以下のような攻撃的ジャミング手段として使えてしまう:
PDFやMarkdownに見えない形で1万字のゴミデータを埋め込む
結果、モデルが誤動作・応答不能・軽量モデルに切替
→ 意図的に相手のAI作業を妨害できる(=AIジャミング)
❗これはOpenAIの「防御側としての脆弱性」でもある
🔚 結論と要求
使用回数とモデル状態を明示せよ
モデル切替は必ず通知せよ
GPT-4.5の実質トークン処理能力を現実に合う形で表示せよ
PDF・長文耐性の改善を速やかに行うべき
OpenAIからの回答
OpenAI
• Just now
Thank you for detailing your feedback and suggestions! You've made valuable points about transparency, usability, and reliability when using the GPT-4.5 model and other OpenAI features. Here’s a summary of how your requests align with current product functionality and what you can do with OpenAI products today: Display Remaining Usage for GPT-4.5
Currently, the interface does not show the remaining number of GPT-4.5 messages for Plus users. Your suggestion to make this visible is highly practical for user experience and has been noted by OpenAI for improvements. Right now, users must estimate their usage or wait for a quota-over message. Notification on Automatic Model Switches
OpenAI does not currently issue an explicit alert when usage limits are reached and another model (like GPT-4o mini) is used automatically. Your concern about clear disclosure and service transparency is important and will be shared as feedback. For now, users must watch for subtle hints, such as model name changes in the dropdown list. Session Usage Report
There is no built-in feature providing an end-of-session or daily usage summary in the UI. Your idea for an automated usage report covering each model, quota, and switch events would indeed benefit professional and transparency-minded users. Usage summaries at this level are not yet part of the product, but this feedback is valuable. Clarify Token Limits for GPT-4.5
While GPT-4.5 advertises support for up to 128,000 tokens, actual capacity can be lower, especially in the web UI. The context window varies by input/output length, and limits may result in errors or model downgrades for long requests. Clearer documentation or practical limits in the UI would enhance trust; this is a common area of improvement requested by users. Mitigate AI Jamming via Hidden Long Text
You're correct: large or hidden input (like invisible text in Markdown or PDFs) can slow performance, trigger downgrades, or cause context issues. Improving detection and prevention of such abuse is actively being explored to maintain quality and reliability. Next Steps for Your Feedback:
All your points are valuable and address key aspects of the OpenAI user experience. Your feedback will be relayed for consideration in future updates. If you have suggestions for additional features or need help navigating current limitations, let me know—I'm here to help you get the most out of OpenAI products! Would you like guidance on maximizing your usage under the current system, or do you have further input or priorities?
✅ OpenAIが公式に認めた内容(要約)
指摘項目 | OpenAIの回答 | 含意 |
---|---|---|
GPT-4.5の残回数が見えない件 | 現在UI上では非表示。改善要望として受理済 | 要望は正当であり、社内でも「必要」と認識 |
モデル自動切替の通知がない件 | 現状は通知なし。 | 「重要なフィードバックとして共有」 |
セッション使用レポート提案 | 未実装だが価値ある提案。プロ向けには有用との認識 | 実装候補入りの可能性あり。影響大 |
128kトークン限界の実質的低さ | 実際の限界はUIや長文構造により低下。ユーザーの共通要望 | 内部でも認識されている弱点。改善候補として共有済 |
AIジャミングリスク | 正当な指摘。対策検討中 | セキュリティ視点でも重要問題。反映の可能性高し |
きっかけ
新しいChatGpt-4.5を試したところ、いきなり残り6回と表示された。使用した記憶がないので、いろいろ調べたところ、回数制限があり、さらに自動的にモデルを切り替えているということが分かった。しかもなんの表示もなく下位モデルに移行してしまうし、コードの場合は逆に回数制限があるモデルに移行するという挙動を取って追いㇽ。これがユーザーにはほぼ気配でしかわからない。
また4.5はコードに特化しているが、入出力の文字数が実は少ないことがわかった。
そして、このような状況だと一気に大量のテキストを送信するとAIの挙動がおかしくなることが分かった。
🔄 回数制限
ChatGPT Plusプラン(月額20ドル)をご利用の場合、GPT-4.5の利用回数は週あたり50回に制限されています。この制限は、最初のメッセージを送信した時点から7日間の期間でカウントされます。たとえば、月曜日の午後3時に最初のメッセージを送信した場合、翌週の月曜日午後3時にリセットされます。現在の使用状況やリセット日時は、モデル選択ドロップダウンでGPT-4.5を選択すると確認できます。
🔄 自動的なモデル切り替えの主な要因
1. 使用回数の制限
-
GPT-4oの使用制限:
- ChatGPT Plusプラン(月額20ドル)では、GPT-4oの使用が3時間あたり最大80メッセージに制限されています。
- この制限に達すると、通知が表示され、引き続き利用するには他のモデル(例:GPT-4o mini)への切り替えが促されます。
-
GPT-4.5の使用制限:
- GPT-4.5は一部のPlusユーザーに提供されており、週あたり約50回の使用制限があります。
- この制限に達すると、他のモデルへの切り替えが行われることがあります。
2. モデルの自動切り替え
- 使用制限に達した場合、ChatGPTは自動的にGPT-4o miniなどの軽量モデルに切り替えることがあります。
- この切り替えは、ユーザーが明示的にモデルを変更しなくても行われることがあります。
対応策がない
自分で使っていればこのように記録することも考えられる。
日付 | 時刻 | 内容概要 | 使用回数 | 備考 |
---|---|---|---|---|
5/11 | 10:45 | コード修正依頼① | 1回目 | 特に問題なし |
GPT-4.5が自動的に他のモデル(GPT-4o miniなど)に切り替わった場合、その事実はUI上で明示されないことが多いため、正確な使用回数の記録が非常に難しいのが実態です。
❗ なぜ記録が難しいのか?
問題点 | 説明 |
---|---|
自動切替が無通知 | 使用制限に達すると、ユーザーに何も知らせずに軽量モデルに切り替わることがある。 |
応答の差がわかりにくい | GPT-4.5とGPT-4o miniの回答の差は微妙なこともあり、気づきにくい。 |
ドロップダウンも信用できない | 表示が「GPT-4.5」でも、裏で切り替わっている場合がある。 |
✅実用的な「劣化判定ルール」と「自動切替検出チェックリスト」
ChatGPT PlusユーザーがGPT-4.5とGPT-4o miniなどのモデル切り替えを実地で推測・記録するための基準です。
✅ 劣化判定ルール(客観的チェック基準)
項目 | 劣化の兆候 | 客観的判断ポイント |
---|---|---|
1. 応答の簡略化 | 詳細な説明がなくなり、1~2文程度で回答される | 例:「このコードは適切です。」のみなど。文の長さが激減 |
2. コード生成の質 | 構文ミス、浅いロジック、誤った前提 | 例:未定義変数の使用、コメントなし、構造ミスが増加 |
3. コンテキスト保持力 | 直前の指示やコードを無視した応答が返る | 「この変数どこ?」などの再質問に的確に答えられない |
4. 推論の深さ | 単純な事実説明しかせず、因果関係・意図を読み取らなくなる | 例:「なぜこうなる?」に対して「それはそうだから」的な説明 |
5. 言語表現の精度 | 表現が稚拙になったり、曖昧な助動詞(たぶん・かもしれません)多用 | 「これは~と思われます」など語調が弱くなる傾向 |
6. 応答速度 | やや早くなる(ただし通信環境にも依存) | GPT-4.5は重く、miniは軽いためレスポンス時間が異なることがある |
自動切替検出チェックリスト
チェック項目 | YES / NO | コメント・記録欄 |
---|---|---|
❶ 長文・構造的な説明が急に出なくなったか? | ☐ / ☐ | |
❷ 明らかな文法ミス・構文ミスが出始めたか? | ☐ / ☐ | |
❸ 過去の指示(前メッセージ)に反応しなくなったか? | ☐ / ☐ | |
❹ 「続きをお願い」と言っても反応が続かないか? | ☐ / ☐ | |
❺ 回答が急にあいまいになったか?(たぶん、かもしれません) | ☐ / ☐ | |
❻ モデル選択はGPT-4.5のままだが、内容が明らかに低品質か? | ☐ / ☐ | |
❼ モデル選択の挙動が不安定だった(クリックしても変化なし) | ☐ / ☐ | |
合計 YES 数(目安) | ☐ / 7 | 3つ以上で切替の可能性大 |
🎯 判定基準の目安
- YESが3つ以上:高確率でGPT-4.5→GPT-4o mini等へ自動切替済み
- YESが5つ以上:ほぼ確定と見て良い。以後の使用は「低精度モード」とみなすべき
さらにこれは意図的に起こせる
慶応の、学生に書かせるレポートで課題PDFを直接ChatGPTに読ませると無関係の出力が出るようにハックした実験、の是非に関する報道見ました
— ひらりん@星野源ファミリー (@maco2171) May 1, 2025
そこの調査では日本の約半数の大学生がChatGPTを使っていないそう
GPT以外のAIを使っている人がいるとしても約半数しかGPT使っていないのは少なくて驚き
慶應のAI対策ガチおもろい
— 陰陽痔 (@onmyoupiles) April 21, 2025
してやられた pic.twitter.com/KkQCHzcKMT
これは意図的にハッキングが可能で、AIに嘘を言わせることができることがわかる。
電子契約書終了のお知らせ
現状ではこの透明度100の文字で契約書を書いてしまえば、契約が成立したことになってしまう。この技術は学生の不正どころかDXの基幹である電子契約の信頼性、PDFの信頼性を破壊している。さらにAIでそれを判定させようとしても妨害ができる可能性まであるわけである。OpenAIの回答もユーザーから同内容の回答が来ているほか、この脆弱性をすでに把握しているものと思われる。