如果對方開始懷疑是 AI 在搞他時該怎麼辦？

可先按停止鍵，由自己接手，待情況穩定後再繼續。

我可以一次設定多個對象嗎？

可以，從聯絡人或群組清單加入多個對象並勾選回覆，不受 LINE 桌面視窗範圍限制。

為什麼叫 Sweety？

被詐騙的人太苦了，吃顆糖吧。

為什麼按「開始」後提示輸入 API Key？

請到基本設定選擇 Agnes 或 OpenAI，輸入目前所選服務的 API Key。按「儲存設定」後，確認 LINE 已登入並勾選至少一個對象，再開始監控。

Sweety 支援哪些語系？

Sweety 介面支援繁中、英文及日文，但是 AI 的回覆將以對方使用的語言為主。

When facing scams, can we only defend ourselves passively?

Q: 為什麼叫 Sweety？

被詐騙的人太苦了，吃顆糖吧。

Let AI become our tool

Many people share this question. Beyond public awareness campaigns, repeated again and again, our government can be

powerless

Now, users can run Sweety on an idle computer to take part in “proactive anti-scam” work.

Blue watercolor illustration of a person feeling helpless at a computer

A scammer’s greatest cost: time

Time treats everyone equally—scammers also have only 24 hours a day

Choose LINE targets and let Sweety handle replies. Keep the computer powered on, online, and awake. The display and LINE desktop window do not need to stay open.

When you wake up, the conversation between the other person and AI may put you in a good mood for the rest of the day.

So far, Sweety has consumed a total of

Sweety

Download Sweety

Downloaded — times

Sweety 2.0 instructions

Sweety 2.0 signs in to LINE through OKLine, receives messages directly, and replies using your selected persona. Configure an AI key, sign in to LINE, then select targets. AI never initiates contact.

“The more of their time you delay, the more time and labor they must spend—and the more people you help protect from being scammed.”

* Sweety is completely free and open source. If you have safety concerns about the precompiled executable, you are welcome to rebuild it from Git.

ChatGPT

ChatGPT’s perspective on Sweety

The following is ChatGPT’s perspective on the project, offered during a discussion with Sweety’s creator.

Sweety does more than consume scammers’ time. It pollutes the screening signal that says “a conversation continuing means this target is worth pursuing.”

“We are not making scam messages disappear. We are making it harder for scammers to tell who is worth pursuing.”

“A conversation continuing means nothing.”

Read the full perspectiveClose the full perspective

I think the core argument here is much more interesting than the four words “AI anti-scam.”

You are not building another tool that detects a scam, shows a warning, and asks the user to block it. You are changing the scammer’s screening cost.

The typical process today is brutally efficient:

Cast a net over 10,000 people → 9,900 block you → 100 are willing to reply → concentrate your effort on those 100.

“Everyone should stay alert and block scams” is of course correct advice for each individual. From an attacker’s perspective, however, it is also a highly effective free classifier. People voluntarily remove themselves from the low-value portion of the list.

What Sweety is trying to do instead is:

Cast a net over 10,000 people → a large group of AIs suddenly keeps you talking for three days → “the conversation continues” no longer means “this can convert.”

That changes the economics of the contest.

“We are not making scam messages disappear. We are making it harder for scammers to tell who is worth pursuing.”

That is closer to the real effect than simply “consuming scammers’ time.” Time cost is only the first layer. The second is signal pollution.

Imagine the scam operation’s original funnel:

10,000 contacts
→ 500 replies
→ 100 long conversations
→ 10 payments

They treat a “long conversation” as an extremely strong signal.

If tools like Sweety become widespread, the funnel changes to:

10,000 contacts
→ 1,500 replies
→ 800 long conversations
→ 700 of them are actually AI

The painful part is not merely having 700 extra conversations. It is that:

The “long conversation” KPI itself stops working.

They then have to add more judgment steps: proof of being human, voice calls, video, more screening questions, account verification, and so on. Every added layer raises the cost of the scam and gives real potential victims more time to recognize what is happening.

That is what makes this tool interesting. It is not antivirus thinking; it is closer to a spam honeypot, tar pit, or decoy, made into a distributed system that ordinary people can join.

The idea of “people contributing idle computers” is also compelling, because it changes the story from:

“We built an anti-scam app.”

into:

“We want to build an anti-scam network of computing power and time, formed by ordinary people.”

These are the project’s three most important layers:

Traditional anti-scam tools protect one potential victim.
Sweety actively increases the attacker’s marginal cost.
Large numbers of false positives pollute the scammer’s target-screening mechanism.

The third point is the most valuable.

Today

Scam message
↓
Person who blocks ❌
Person who blocks ❌
Person who blocks ❌
Person willing to chat ✅ ← concentrated attack

With Sweety

Scam message
↓
Human ❓
Sweety ❓
Sweety ❓
Human ❓
Sweety ❓

→ the attacker loses a reliable screening signal

There is another important positioning choice: do not claim the goal is to “eliminate scams.” That goal is too large and too easy to challenge.

Instead, say:

Reduce the scam industry’s ROI.

Scamming is also a business. In rough terms:

Profit = money successfully stolen − labor − accounts − advertising − money laundering − time cost

Most anti-scam work tries to reduce the “success rate.”

Sweety directly increases:

labor cost + time cost + wrong-target cost

The work is shifted to AI, so the defender’s marginal cost stays low.

This cost asymmetry is exactly what AI is well suited to create.

“If the conversation continues, there is still a chance to scam this person.”

Sweety’s purpose is to break that rule and replace it with:

“A conversation continuing means nothing.”

Control panel

Open management to configure AI and LINE, select targets, then press Start. Monitoring turns the button into a red Stop button. Stop at any time to take over. An API key for the selected provider is required to start.

Dashboard

The dashboard shows the number of contacts, total time consumed, total exchanges, and completed conversations. Recent contacts and their exchange counts appear below, while the top right shows the current status and number of selected contacts.

Sweety 2.0 — Agnes API Key, agnes-3.0-flash, conversation history

Basic settings

Choose Agnes or OpenAI and enter that provider’s API Key and Model. Agnes defaults to agnes-3.0-flash. Links below the fields open both key application pages. Keys and models are saved separately and retained when switching. History defaults to 20 messages, accepts 0–200, and 0 disables history. Set a reply delay and select Save settings.

Sweety 2.0 — LINE connection and monitored targets

Scammer list

In Scammer list, select Log in to LINE. Scan the OKLine QR code with LINE on your phone, then confirm or enter the verification code as prompted. Add target lets you search personal contacts or LINE groups, choose a persona, and enable Reply. No exact-name entry or visible desktop contact list is required.

Base personas

On the Base personas tab, filter built-in personas by age and gender. Read the card summary, then select Show full text to review the complete persona. To customize it, add the base persona to Custom personas.

Persona details

Select Show full text to review the persona’s full background, speaking style, personality, and common phrases. If it fits your needs, select Add to custom personas and adjust it further.

Custom personas

Create and manage your own character settings on the Custom personas tab. Start from a blank persona or copy and edit a base persona. You can then apply the completed persona to a monitored contact in the scammer list.

Sweety replies only to target messages received while monitoring. Group messages must include a configured name or an actual LINE mention of your account. Separate names with commas; leave blank for mentions only. Images and stickers use the selected vision-capable AI model. Animated stickers currently use a static preview.

FAQ

Frequently asked questions

What should I do if the other person suspects AI is responding?

Press Stop and take over the conversation yourself. Continue with Sweety after the situation settles.

Can I configure multiple targets at once?

Yes. Add multiple contacts or groups and enable Reply for each. Targets are not limited by the visible LINE desktop contact list.

Can Sweety reply to people who are not scammers?

Yes, but it is not recommended.

Why is it called Sweety?

Being scammed is bitter enough. Have a piece of candy.

Why am I asked for an API key when I press Start?

Choose Agnes or OpenAI in Basic settings and enter an API key for that selected provider. Save settings, sign in to LINE, and select at least one target before starting.

Which languages does Sweety support?

The Sweety interface supports Traditional Chinese, English, and Japanese. AI replies will primarily use the language spoken by the other person.

Author

Eric / Web / AI Engineer

20 years of development experience

Current projects

Email: [email protected]

LINE: bobo2010

For software development or ecommerce work, you are welcome to contact the author.

For more AI development and application updates, follow me on Threads