================================================
To:
Shri Ashwini Vaishnaw, Hon'ble Union Minister of Electronics & Information Technology
Subject:
An impartial, AI-assisted "Mute the Mike" system for the Lok Sabha: platform,
architecture and agent workflow
================================================
Respected Shri Ashwini Vaishnawji,
Today's report that the Government has asked TV channels to monitor debates and mute panellists whose remarks turn abusive prompted me to write Freedom of Speech ? Yes : Freedom to Hate ? No. In it I revisited two earlier suggestions of mine: that the Hon'ble Speaker should be able to simply mute the mike of a member who uses unparliamentary language (Virtual Parliament, June 2020), and a technology framework that rates speech on a 1 to 5 scale and auto-mutes at levels 3 to 5 (Hate Speech: Prevention Mechanism, December 2023).
The obvious objection is: who decides what is hateful? If an anchor decides, or a ruling-side official decides, the system will be seen as partisan, however good its intentions. So in this follow-up I set out how such a system can be built so that it is not only fair, but visibly and verifiably fair to every party, every language and every member. MeitY's existing assets, especially Bhashini and the IndiaAI Mission, make India unusually well placed to build it.
Seven principles that make it impartial
- The House owns the rules, not the Government. Under Article 118 the Lok Sabha regulates its own procedure. The rulebook, word lists and severity scale should be framed by the Rules Committee with members from all parties, and adopted by the House. MeitY's role is only to supply the technology.
- Published rules, not secret judgement. Every rule, every listed expression and every severity definition is public, versioned and dated. A member can know in advance exactly what will trigger a mute.
- Same rule for everyone. The system does not know, and is not allowed to know, a member's party when it scores speech. Party affiliation is used only afterwards, to test for bias.
- AI recommends; the Chair decides. Only the most extreme category (open incitement to violence) is muted automatically, and even that can be reversed by the Chair in one touch.
- Mute is short and reversible. A mute lasts seconds, not the rest of the debate, and the member is told why on their console.
- Everything is recorded. Every score, every action and every override goes into a tamper-evident audit log.
- Fail open. If any component fails, microphones stay on. A technical fault must never silence a member.
Technology platform and system architecture
The platform sits between the members' microphones and the sound and broadcast system, entirely inside Parliament's own secure network. It has six layers.
Layer | What it does | Built on |
|---|---|---|
1. Capture | Takes a digital tap of each seat microphone. Identifies the speaker from the seat console login, not from voice or face. | Existing Lok Sabha conference system |
2. Language | Streams speech to text in all 22 scheduled languages, including code-mixed speech such as Hinglish, within about a second. | Bhashini speech and translation models |
3. Understanding | A team of AI agents (next section) checks context and scores each utterance against the published 1 to 5 scale. | Indian models trained under the IndiaAI Mission, hosted on-premises |
4. Policy | A plain rules engine, not an AI, converts scores into actions exactly as the House has written them. | Open, versioned rule files |
5. Action | Warns the member, alerts the Chair, or mutes the mike for a fixed few seconds. Also controls a short delay on the Sansad TV feed. | Sound system and broadcast controls |
6. Accountability | Tamper-evident audit log, Chair's dashboard, member's appeal portal and a public transparency dashboard. | NIC infrastructure, CERT-In audited |
The whole stack runs on servers inside Parliament, with no dependence on the internet or any foreign cloud. Its source code, models and test results are open to inspection by members and by an independent technical auditor, for example a consortium of IITs.
The AI agents workflow
No single AI model decides anything. Each utterance passes through a chain of specialised agents, and two independent agents must agree before any action is proposed.
Every path to a mute passes through two independent scores, a rules engine the House wrote, and the Chair; every step lands in the audit log that feeds appeals and public reporting.
- Transcription Agent. Converts speech to text in the original language, and keeps the audio for the record.
- Context Agent. Checks who is being addressed and how: is the member quoting someone, reading from a document, reporting a threat they received, or being sarcastic? Most false alarms come from missing this.
- Lexicon Agent. Matches the text against the Lok Sabha Secretariat's own published list of unparliamentary expressions. This step is fully deterministic and explainable.
- Primary Scoring Agent. Rates the utterance from 1 to 5 against the published rubric, and must cite the exact words and the rule it relied on.
- Second Opinion Agent. A model from a different developer and training lineage scores the same utterance independently, actively looking for innocent readings.
- Policy Engine. Not an AI. It applies the House's written rules to the two scores. If the two agents disagree, the lower score prevails and the case goes to the Chair.
- Action Agent. Carries out only what the Policy Engine allows: a warning, an alert, or a timed mute.
- Audit Agent. Writes every step, with timestamps, scores, reasons and rule version, into the tamper-evident log.
- Fairness Monitor Agent. Runs in the background, comparing outcomes across parties, languages, regions and genders, and raises an alarm if any group is treated differently.
What happens at each level
Level | Meaning | Action |
|---|---|---|
1 | Distasteful | Logged only |
2 | Offensive or unparliamentary | Private reminder on member's console; Chair may expunge |
3 | Personal attack or abuse of a community | Chair alerted with transcript; Chair decides to mute |
4 | Dehumanising or threatening language | Chair alerted; automatic 10-second mute unless Chair cancels within 3 seconds |
5 | Inciting violence or outright hate | Immediate automatic mute; Chair can restore at once |
The exact thresholds and mute durations above are illustrative. They are for the Rules Committee to set.
Safeguards that prove impartiality
Human oversight. The Hon'ble Speaker, or whoever is in the Chair, keeps final authority at every level. The Chair's dashboard shows the transcript, both agents' scores and the rule cited, with one-touch buttons to mute, restore or dismiss. Every override is logged, so the Chair is accountable too.
Tamper-evident audit trail. Each log entry is cryptographically chained to the previous one, so no entry can be quietly altered or deleted. Copies of the daily log fingerprint can be held by the Secretary-General and by the leaders of all recognised parties.
Right of appeal. A muted member can appeal through a portal within 24 hours. A bipartisan panel reviews the audio and the log. If the mute was wrong, the words are restored to the record and the correction is announced in the House.
Bias testing before and during use.
- Mirror tests: thousands of identical statements are run with only the party, religion, caste, region or speaker's gender swapped. The scores must match.
- Language parity: the same statement in all 22 languages must get the same score. A Tamil or Bengali speaker must not be muted more often than a Hindi or English speaker for the same words.
- Outcome parity: monthly reports compare warnings, mutes and overturned mutes by party and language. Any gap beyond an agreed limit suspends automatic muting until fixed.
- Independent red team: researchers nominated by different parties try to fool the system or expose bias, and their findings are published.
Public transparency. A monthly public dashboard shows how many alerts, mutes and successful appeals occurred, broken down by party and language, along with the current rulebook version.
A cautious, phased rollout
- Consultation. The Rules Committee and an all-party group agree the rulebook, the severity scale and the fairness limits.
- Shadow mode, one full session. The system listens, transcribes and scores, but takes no action. Its private reports are compared with the Chair's own rulings, and the bias tests are published.
- Advisory mode. The Chair receives alerts and decides every action personally.
- Limited automatic mute. Only level 5 is muted automatically, and only after the shadow and advisory results pass independent audit.
- Wider use. Once proven in the Lok Sabha, the same platform can serve the Rajya Sabha, State Assemblies, and the TV debate channels mentioned in today's report, giving all of them one consistent, published standard instead of each anchor's personal judgement.
Sir, I recognise that this will take time and broad political agreement. But a system designed this way would protect free debate rather than threaten it, because no member, of any party, could be silenced for anything other than what the House itself has written down in advance.
I would be honoured to discuss this further with your officers at MeitY, Bhashini or the IndiaAI Mission.
With regards,
Hemen Parekh
myblogepage.blogspot.com · hcpblogs@gmail.com
No comments:
Post a Comment