Scam Call Filtering: Billions Flagged, Little Independent Proof of What AI Adds

The use case

In September 2024 Bharti Airtel, an Indian mobile operator, announced a network filter that, by its own account, flags 100 million suspected spam calls and 3 million spam text messages every day (operator-reported, at launch on 25 September 2024)¹. The filter sits inside the operator's network, so it screens traffic before it reaches a handset. It is one example of a wider group of tools that block or label suspected scam calls and texts.

What counts as AI here. This paper counts only systems that are stated to use machine learning or comparable AI methods, meaning models that learn patterns, such as unusual calling behaviour, from past data. Several older techniques do similar work and are not credited as AI: blocklists of known bad numbers, call-authentication standards, do-not-originate lists (numbers that should never place calls), limits on messages per SIM card, and reputation lists built from customer reports. Ofcom's UK measures, for example, are described as volume limits, number blocking and detection of scam links and numbers based on customer and third-party reports, with no AI method stated¹⁰. Most blocking counts published by regulators and operators do not say which method produced them. Bank-side payment fraud detection and AI-generated voice scams are outside this paper's scope.

Where it is running

In the United States, regulator rules since 2020 allow carriers to block calls based on reasonable analytics designed to identify unwanted calls, generally through a third-party analytics company partner¹²; the sources read give no AI statement. Earlier tests exist: a vendor-designed comparison of three US spam-detection providers was run in December 2018¹⁴.

Counts are not comparable. Airtel counts flagged calls, Telstra and the Australian regulator count blocked calls, and Airtel's customer alerts are warnings rather than blocks¹²⁷⁸. The figures cannot be added together or ranked.

What the evidence shows

Offline accuracy. The strongest published result is academic. A peer-reviewed study analysed anonymised call detail records (the logs a network keeps of each call), more than 1 billion records from one large telecom provider over 10 days. Its caller trust-score features combined with machine learning models reached a true-positive rate of around 97% with a false-positive rate below 0.01% (abstract; measured offline on one operator's data, not in a live deployment)³. Because the data came from one operator over 10 days, the result shows what is possible on that sample; it does not show how the model would perform as scam tactics change.

Operator-reported effect. Airtel reported that it alerted more than 252 million unique customers about suspicious calls, "leading to a 12% decline in response to such calls" (operator-reported, to early December 2024; how the decline was measured is not explained)². A fall in people answering is not the same as a fall in losses.

Vendor-reported effect. Google reported that its AI-powered detection found 20 times more deceptive pages for removal from Search results, and that airline-impersonation schemes fell by over 80% in 2024 (vendor-reported, no baseline or method given; these figures concern web search, not calls or texts)⁴.

Older vendor test. In a December 2018 test designed by Hiya, with disputed numbers classified by a contractor, Hiya identified 87% of spam numbers with a 0.25% error rate, against 55% and 1.26% for First Orion and 34% and 1.48% for TNS (vendor-designed and vendor-promoted, Hiya ranked first; data older than three years; no machine learning stated)¹⁴.

Outcomes for consumers. In 2025, phone calls accounted for 11% and texts for 7% of people who reported a scam loss to the US Federal Trade Commission, by contact method (consumer self-reports covering all scams, with or without filtering)¹⁶. No source read shows how much of any change in losses is due to filtering. The published counts measure how many calls and messages were flagged, blocked or alerted, and none of the sources read measures how many scams were prevented.

What it takes: cost, integration, governance

Cost. No public price or cost per unit was found for any deployment.

Data and integration. Airtel states that its filter was developed in-house, analyses patterns such as call and SMS frequency and call duration, flags changes of handset identifier (IMEI), and checks links in messages against a blacklist¹. In the United States, carriers generally work through third-party analytics partners¹².

Rules. The US Federal Communications Commission's Eighth Report and Order (adopted 27 February 2025) requires providers in the call path to block calls from do-not-originate lists. It makes one signalling code (SIP code 603+) the exclusive way to tell callers that analytics-based blocking occurred, to help correct erroneous blocking, and it declines to adopt a mandate for analytics-based blocking⁵. The US Government Accountability Office reported that the rule takes effect on 25 March 2026, except amendments to 47 C.F.R. § 64.1200(o), which are delayed indefinitely pending approval of information-collection requirements⁶. Existing rules already require a dispute point of contact and a status update within 24 hours⁵. In the UK, Ofcom has published a statement and guidance for mobile operators and business messaging aggregators⁹; new rules start on 18 January 2027 for person-to-person messaging and 15 July 2027 for business messaging¹⁰. In India, a regulator directive requiring AI/ML-based spam detection is dated by MediaNama to June 2023². In Australia, telcos report blocking counts to the regulator, which uses them to identify trends⁷.

Where it falls short

Counts without a counterfactual. No source read compares scam losses or scam exposure with and without a filter. The headline counts have no denominator for how many scam attempts were never caught.

Errors. In 2020, providers told the FCC that fewer than 1%, and as few as 0.2%, of blocked calls were false positives (provider-reported)¹³. The article's authors extrapolated that 0.2% of one analytics provider's 2019 volume would be about 1.6 million calls blocked in error, which is their estimate and not a measurement¹³. In 2025 the FCC was still adding measures to correct erroneous blocking⁵. A US survey reported by EMARKETER found that 66% of adults had ignored an important call assuming it was spam, and that 45% of legitimate calls missed after being mistaken for spam were from doctors or medical providers (survey by Cloaked, a vendor; sample size and date not given; the figures concern people's own judgement, not filter errors specifically)¹⁵. In the UK, the trade body Comms Council UK warned of false positives in Ofcom's approach (reported 16 July 2026)¹⁰.

Challenge to the main claims. The most favourable results are the least robust. The 97% detection rate comes from one operator's records over 10 days and was not tested live³. The 12% fall in responses is operator-reported with no control group described². The only head-to-head accuracy test was designed by the vendor that won it, in 2018¹⁴. Much of the blocking also happens alongside non-AI levers: the Australian regulator issued 945 compliance alerts to businesses in one quarter⁷, and Ofcom's measures rely on rules and customer reports¹⁰. Gains therefore cannot be credited to AI alone.

Who benefits. Airtel reported that customers aged 36 to 60 received 48% of spam calls and senior citizens about 8% (operator-reported, 2024); this shows who receives spam, not who is protected². No source read measures uneven effects across groups.

Withdrawals and challenges. No withdrawn, paused or legally challenged scam-filtering programme was found in the sources read. The search for these was limited, so this is not evidence that none exist.

Phase: In early use — why

The hypothesis was In wide use. It is changed to In early use. Only deployments confirmed to use AI count, and these are Airtel's filter (operator-reported, 2024) and Google's on-device detection in Messages (vendor-reported, 2025)¹⁴, plus one academic study³. The large blocking counts in Australia, the UK and the United States do not state the method⁷⁸¹⁰¹². The share of genuine AI use across the sector cannot be established, so the lower phase is chosen.

What to watch next

• Ofcom's final decision on calls from abroad that spoof UK mobile numbers, planned for "FY Q2 2026" in a letter of 8 July 2026¹¹.
• Ofcom's UK messaging rules, starting 18 January 2027 and 15 July 2027¹⁰.
• A Federal Register notice of OMB approval for the US rules, which starts the 90-day clock for do-not-originate blocking⁵⁶.
• Any FCC order following its 2023 inquiry into whether AI-based tools can enhance blocking and whether blocking decisions can be independently substantiated¹⁷.

Questions to ask before adopting

• Does the supplier state whether a machine-learning model, a blocklist or a rule set makes each blocking decision?
• What share of flagged calls are blocked, labelled or only alerted, and what is the measured rate of lawful calls wrongly affected?
• Has anyone outside the supplier tested the system against a control group or an unseen dataset?
• How quickly, and at what cost, can a wrongly labelled caller or recipient get a correction?
• Which privacy, redress and notification rules apply to the call data in each market served?