Fact-Checking Policy

Purpose

Fact-checking at DeshiAI is the structured process of determining whether published claims are supported by reliable evidence, correctly attributed, appropriately qualified and consistent across the headline, article, images, captions, metadata and disclosures. Fact-checking is a substantive editorial process, not a decorative label. DeshiAI does not describe material as fact-checked unless meaningful verification has taken place. Authors remain responsible for the accuracy of their submissions, while editorial review provides an additional layer of scrutiny rather than shifting all responsibility to an editor.

Proportionate Verification

Different claims require different levels of verification. The intensity of checking depends on the importance of the claim, potential harm, public relevance, reliability of the source, degree of dispute, whether a private individual is involved, whether wrongdoing is alleged, whether the subject concerns health, finance, elections, security or safety, whether anonymous sources are used, whether media may have been manipulated or synthetically generated, and whether an error could create significant legal, technical or reputational consequences. High-risk claims require stronger evidence, clearer attribution and more complete records than routine background information.

Author and Editor Responsibilities

Before submission, authors identify the evidence supporting material claims, retain relevant documents and notes, distinguish firsthand knowledge from secondhand information, verify names, dates, organizations, products and technical specifications, check quotations, disclose uncertainty, identify interested-party sources, declare material conflicts, disclose significant AI or automated-tool use, preserve important calculations and alert the editor to legal, privacy, security or safety concerns. Difficulty finding supporting evidence does not justify presenting an uncertain claim as established fact.

Editors independently test important claims according to their significance. This may involve reviewing primary documents, opening cited sources, checking recordings, contacting organizations, comparing independent reporting, reproducing calculations, examining images and metadata, checking technical documentation, consulting qualified specialists, verifying chronology and location, contacting affected parties and documenting unresolved uncertainty. A claim is not accepted simply because it comes from an experienced writer, recognized company, technical expert or prominent institution.

Claim Classification

DeshiAI distinguishes verified facts from allegations, estimates, inferences, predictions, opinions, quotations and official findings. A verified fact is supported by reliable evidence within the limits of that evidence. An allegation is a claim that has not been finally established. An estimate is an approximation whose methodology and uncertainty are relevant. An inference is a conclusion drawn from established information and identified as such. A prediction concerns a future result and should identify meaningful assumptions. Opinion expresses judgment or argument while its factual premises remain checkable. A quotation establishes that someone made a statement, not that the statement itself is true. An official finding is described according to the authority that issued it, its jurisdiction, procedural stage and appeal status.

Source Evaluation

Sources are assessed according to identity, expertise, proximity to the underlying information, basis of knowledge, motivation, interests, independence, consistency, documentary support, previous reliability and ability to be independently corroborated. An official source may reliably establish what an institution announced or decided without establishing that every factual assertion accompanying the decision is correct. A technology company may be authoritative about what its own product does or what it claims to have launched, while remaining an interested party regarding performance, safety, market position or social impact.

DeshiAI prefers primary material where available, including official documentation, regulations, court decisions, research papers, original datasets, technical specifications, company filings, recorded statements, direct interviews, product documentation and firsthand observation. Secondary reporting can provide useful context and leads. When another publication conducted the original reporting, DeshiAI attributes that work rather than presenting it as independently discovered information.

Quotations, Translation and Identity

Quotations are checked against recordings, original transcripts, contemporaneous notes, official written statements, reliable published material or direct confirmation. They may be shortened where the omission does not change meaning. Words should not be rearranged to manufacture a stronger or more dramatic statement. Where audio or source material is unclear, the uncertainty should be resolved where possible or the material should be paraphrased.

Translated material should preserve meaning, relevant tone and technical context. Editors consider whether specialist human review is necessary, particularly for legal, scientific or technical terminology, whether the quotation is disputed and whether automated translation was used. Significant translation limitations are disclosed where relevant. Names are checked for spelling, preferred form, title, current or former role, organization and the risk of confusing one person, company or product with another.

Dates, Chronology and Location

Dates are checked against original records wherever reasonably possible. Editors verify event dates, announcement dates, publication dates, product-release dates, effective dates, relevant time zones and the sequence in which events occurred. They also consider whether a source existed at the time being described and whether a later update has altered the apparent timeline. Chronological inconsistencies can expose mistaken identity, recycled information, manipulated media or unsupported claims of causation. Location claims are checked through reliable records, visual evidence, official information or other appropriate methods when location is material to the story.

Numbers and Calculations

Numerical claims are checked for their source, unit, currency, period, numerator, denominator, sample size, methodology, rounding and relevant assumptions. Particular care is taken with percentages, percentage points, averages, totals, revenue, profit, valuation, downloads, registered users, active users, model parameters, benchmark scores, latency, accuracy rates, energy consumption and other technology-specific measurements. Comparisons should use compatible definitions and testing conditions. Calculations should be reproducible from the stated method, and assumptions should be visible when they materially affect the result.

Statistics, Research and Polling

Research coverage identifies the authors, institution, funding, publication status, peer-review status, study design, sample, population, geography, period, uncertainty, limitations, competing evidence and relevant replication. A preprint is not presented as established peer-reviewed evidence. Statistical significance does not by itself demonstrate practical importance, and findings from a narrow population or controlled environment are not presented as universal.

Survey and polling coverage identifies the commissioner, conducting organization, field dates, sampling method, target population, sample size, question wording, response options, weighting, margin of error where applicable, non-response, geography and relevant limitations. A voluntary online poll or social-media survey is not described as representative of a wider population without supporting methodology.

Company, Product and Market Claims

Technology companies and product providers may make claims concerning users, downloads, revenue, market share, model capability, benchmark performance, safety, security, partnerships, certifications, environmental impact, adoption or regulatory status. DeshiAI identifies whether a claim is independently audited, externally tested, officially certified, internally measured, self-reported, projected, based on supplied documentation or not independently verifiable. Marketing descriptions such as fastest, leading, revolutionary, most secure, most accurate, largest, first or industry-changing are not adopted as factual conclusions without sufficient evidence.

Legal Claims

Legal reporting identifies the relevant jurisdiction, legal authority, procedural stage and whether a matter is civil, criminal, regulatory or administrative. Editors distinguish interim orders from final decisions, allegations from findings, charges from convictions and investigations from established wrongdoing. A claim made in a lawsuit is not itself a judicial finding. Where legal terminology differs between jurisdictions, the article explains the distinction rather than assuming equivalent meanings. Qualified legal review may be obtained when interpretation is complex, uncertain or consequential.

Images, Screenshots, Video and Social Media

Visual material is checked for original source, uploader, date, location, context, metadata, editing, synthetic generation, prior publication and consistency with accompanying claims. Screenshots are examined for possible cropping, altered interfaces, missing context, fabricated accounts and outdated information. A genuine image can still be misleading when it is old, unrelated to the event or incorrectly described. Detection tools may assist verification but are not treated as conclusive evidence on their own.

Social posts are assessed for account identity, publication date, edits or deletion, archived versions, context, impersonation, parody, translation, media authenticity and whether the account has direct knowledge of the subject. A verification badge, follower count or viral distribution does not establish that a claim is true.

Anonymous Sources and Right of Reply

Claims based on anonymous sources receive additional scrutiny. The responsible editor should know the source’s identity, basis of knowledge, motivation, relevant conflicts, risk of retaliation, corroborating evidence and what information can safely be disclosed. A single anonymous source supports publication only in exceptional circumstances and does not by itself justify unsupported accusations or personal attacks.

Contacting a criticized person, company or organization is also an important part of verification. A response may identify factual errors, missing records, different chronology, legal developments, technical limitations, mistaken identity, contextual omissions or relevant admissions. A response is not automatically accepted as true, but it is relevant evidence and should be assessed rather than reproduced without scrutiny.

Artificial Intelligence Verification

AI output is not treated as a source simply because it is fluent, detailed or confident. AI-generated facts, citations, quotations, case references, statutes, biographies, job titles, calculations, research summaries, translations, technical explanations and image descriptions must be checked against reliable underlying evidence. Authors and editors remain alert to fabricated citations, invented organizations, conflated individuals, outdated information, incorrect product specifications, synthetic quotations, numerical mistakes and plausible-sounding explanations that lack support.

High-Risk Claims and Headline Review

Enhanced verification applies to allegations of crime or misconduct, fraud, abuse, medical danger, investment claims, election information, national security, cybersecurity incidents, threats, confidential documents, leaked datasets, children, sexual material, suicide, private individuals, discrimination, public-safety warnings and claims likely to cause significant reputational or financial harm. High-risk work should maintain a claim-by-claim evidence record or an equivalent verification trail.

The headline is checked separately from the article. Editors ask whether its wording and implications are supported, whether an allegation has been turned into a fact, whether the scale of a technical result is exaggerated, whether a person or organization is correctly identified, whether a material qualification has been omitted, whether causation is implied without evidence, whether a quotation has been distorted, whether urgency is manufactured and whether commercial status is concealed. A factually supported article can still fail editorial standards if its headline creates a materially misleading impression.

Metadata and Final Reconciliation

Fact-checking extends to the SEO title, meta description, URL slug, category, tags, image metadata, author information, publication date, update date, article type, social preview and structured data. Machine-readable information should not contradict the visible article.

Before publication, the editor confirms that unresolved claims have been removed or appropriately qualified, quotations match their sources, calculations and numbers are consistent, names and dates are correct, source links function, disclosures are complete, the headline accurately reflects the article, classification is appropriate, and the final rendered page matches the reviewed version. Any new errors introduced through formatting, templates or publishing systems should be corrected before release.

Uncertainty, Failed Verification and Corrections

Where evidence is incomplete, DeshiAI uses language that accurately reflects the available information, such as according to the company, available records indicate, the estimate is based on, the evidence does not establish, the claim could not be independently confirmed, the sample is limited, the finding remains preliminary, the allegation is disputed or no final decision has been issued. Qualification should clarify genuine uncertainty and must not be used as a formula for publishing an otherwise unsupported claim.

When a material claim cannot be adequately verified, DeshiAI may remove it, qualify it, seek additional evidence, delay publication, narrow the article’s scope, identify it as an allegation where independently newsworthy or decline publication. A deadline does not transform an unverified statement into fact. Evidence discovered after publication may require a correction, clarification, update, editor’s note or retraction depending on its significance.

Contact

Specific factual concerns may be sent to corrections@deshiai.in. The request should identify the article, exact statement, reason for the concern, proposed correction and supporting evidence.